protect sleepq_len access by sleep_q.lock and move
pxmitpriv->lock after sleep_q.lock release.
This fixes and completes a lockdep warning silencing
done in a prevoius commit where accesses to sleep_q
related fields were protected by sleep_q.lock instead
of pxmitpriv->lock.
Note that sleep_q.lock is already taken inside
rtw_free_xmitframe_queue so we just wrap sleepq_len
access.
Moved pxmitpriv->lock after sleep_q.lock release to
avoid locks nesting.
Fixes: 78a1614a81f0 ("staging: rtl8723bs: remove possible deadlock when disconnect")
Reported-by: Hans de Goede <[email protected]>
Tested-by: Fabio Aiuto <[email protected]>
Signed-off-by: Fabio Aiuto <[email protected]>
---
drivers/staging/rtl8723bs/core/rtw_sta_mgt.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c b/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
index bf090f3b1db6..c98918e02afe 100644
--- a/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
+++ b/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
@@ -294,10 +294,12 @@ u32 rtw_free_stainfo(struct adapter *padapter, struct sta_info *psta)
/* list_del_init(&psta->wakeup_list); */
- spin_lock_bh(&pxmitpriv->lock);
-
rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
+ spin_lock_bh(&psta->sleep_q.lock);
psta->sleepq_len = 0;
+ spin_unlock_bh(&psta->sleep_q.lock);
+
+ spin_lock_bh(&pxmitpriv->lock);
/* vo */
/* spin_lock_bh(&(pxmitpriv->vo_pending.lock)); */
--
2.20.1
Hello Hans,
On Mon, Sep 13, 2021 at 03:24:44PM +0200, Hans de Goede wrote:
> Hi Fabio,
>
> > Note that sleep_q.lock is already taken inside
> > rtw_free_xmitframe_queue so we just wrap sleepq_len
> > access.
> >
> > Moved pxmitpriv->lock after sleep_q.lock release to
> > avoid locks nesting.
> > rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
> > + spin_lock_bh(&psta->sleep_q.lock);
>
> AFAICT this needs to be above the rtw_free_xmitframe_queue() ?
as I wrote on the changelog, the sleep_q.lock is already
taken inside rtw_free_xmitframe_queue. If I put the
sleep_q.lock above that function a soft lock occurs when
I disconnect.
So I put it just below rtw_free_xmitframe_queue.
Things works fine this way.
Please tell me if there's a best way to do it.
>
> Regards,
>
> Hans
thank you,
fabio
Hi Fabio,
On 9/13/21 3:03 PM, Fabio Aiuto wrote:
> protect sleepq_len access by sleep_q.lock and move
> pxmitpriv->lock after sleep_q.lock release.
>
> This fixes and completes a lockdep warning silencing
> done in a prevoius commit where accesses to sleep_q
> related fields were protected by sleep_q.lock instead
> of pxmitpriv->lock.
>
> Note that sleep_q.lock is already taken inside
> rtw_free_xmitframe_queue so we just wrap sleepq_len
> access.
>
> Moved pxmitpriv->lock after sleep_q.lock release to
> avoid locks nesting.
>
> Fixes: 78a1614a81f0 ("staging: rtl8723bs: remove possible deadlock when disconnect")
> Reported-by: Hans de Goede <[email protected]>
> Tested-by: Fabio Aiuto <[email protected]>
> Signed-off-by: Fabio Aiuto <[email protected]>
> ---
> drivers/staging/rtl8723bs/core/rtw_sta_mgt.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c b/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
> index bf090f3b1db6..c98918e02afe 100644
> --- a/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
> +++ b/drivers/staging/rtl8723bs/core/rtw_sta_mgt.c
> @@ -294,10 +294,12 @@ u32 rtw_free_stainfo(struct adapter *padapter, struct sta_info *psta)
>
> /* list_del_init(&psta->wakeup_list); */
>
> - spin_lock_bh(&pxmitpriv->lock);
> -
> rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
> + spin_lock_bh(&psta->sleep_q.lock);
AFAICT this needs to be above the rtw_free_xmitframe_queue() ?
Regards,
Hans
> psta->sleepq_len = 0;
> + spin_unlock_bh(&psta->sleep_q.lock);
> +
> + spin_lock_bh(&pxmitpriv->lock);
>
> /* vo */
> /* spin_lock_bh(&(pxmitpriv->vo_pending.lock)); */
>
Hi Fabio,
On 9/13/21 3:39 PM, Fabio Aiuto wrote:
> Hello Hans,
>
> On Mon, Sep 13, 2021 at 03:24:44PM +0200, Hans de Goede wrote:
>> Hi Fabio,
>>
>
>>> Note that sleep_q.lock is already taken inside
>>> rtw_free_xmitframe_queue so we just wrap sleepq_len
>>> access.
>>>
>>> Moved pxmitpriv->lock after sleep_q.lock release to
>>> avoid locks nesting.
>
>>> rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
>>> + spin_lock_bh(&psta->sleep_q.lock);
>>
>> AFAICT this needs to be above the rtw_free_xmitframe_queue() ?
>
> as I wrote on the changelog, the sleep_q.lock is already
> taken inside rtw_free_xmitframe_queue. If I put the
> sleep_q.lock above that function a soft lock occurs when
> I disconnect.
>
> So I put it just below rtw_free_xmitframe_queue.
>
> Things works fine this way.
>
> Please tell me if there's a best way to do it.
Hmm I see, this may work, but the sleepq_len access
really should be protected by the same lock as the freeing
of the queue is without dropping it in between.
That rtw_free_xmitframe_queue() takes the sleep_q.lock
then to me that signals that other (higher-level) functions should
not take sleep_q.lock at all, since this is then private to the
functions operating on the sleep_q.
I've an idea how we we can possibly tackle this, but I'm not sure
yet I will try to make some time to look into this tomorrow or
the day after.
Regards,
Hans
On Mon, Sep 13, 2021 at 05:12:19PM +0200, Hans de Goede wrote:
> Hi Fabio,
>
> On 9/13/21 3:39 PM, Fabio Aiuto wrote:
> > Hello Hans,
> >
> > On Mon, Sep 13, 2021 at 03:24:44PM +0200, Hans de Goede wrote:
> >> Hi Fabio,
> >>
> >
> >>> Note that sleep_q.lock is already taken inside
> >>> rtw_free_xmitframe_queue so we just wrap sleepq_len
> >>> access.
> >>>
> >>> Moved pxmitpriv->lock after sleep_q.lock release to
> >>> avoid locks nesting.
> >
> >>> rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
> >>> + spin_lock_bh(&psta->sleep_q.lock);
> >>
> >> AFAICT this needs to be above the rtw_free_xmitframe_queue() ?
> >
> > as I wrote on the changelog, the sleep_q.lock is already
> > taken inside rtw_free_xmitframe_queue. If I put the
> > sleep_q.lock above that function a soft lock occurs when
> > I disconnect.
> >
> > So I put it just below rtw_free_xmitframe_queue.
> >
> > Things works fine this way.
> >
> > Please tell me if there's a best way to do it.
>
> Hmm I see, this may work, but the sleepq_len access
> really should be protected by the same lock as the freeing
> of the queue is without dropping it in between.
>
> That rtw_free_xmitframe_queue() takes the sleep_q.lock
> then to me that signals that other (higher-level) functions should
> not take sleep_q.lock at all, since this is then private to the
> functions operating on the sleep_q.
>
> I've an idea how we we can possibly tackle this, but I'm not sure
> yet I will try to make some time to look into this tomorrow or
> the day after.
I'm just going to go and revert the original change here until you all
can sort it out :)
thanks,
greg k-h
Hi Greg,
On Fri, Sep 17, 2021 at 04:25:12PM +0200, Greg KH wrote:
> On Mon, Sep 13, 2021 at 05:12:19PM +0200, Hans de Goede wrote:
> > Hi Fabio,
> >
> > On 9/13/21 3:39 PM, Fabio Aiuto wrote:
> > > Hello Hans,
> > >
> > > On Mon, Sep 13, 2021 at 03:24:44PM +0200, Hans de Goede wrote:
> > >> Hi Fabio,
> > >>
> > >
> > >>> Note that sleep_q.lock is already taken inside
> > >>> rtw_free_xmitframe_queue so we just wrap sleepq_len
> > >>> access.
> > >>>
> > >>> Moved pxmitpriv->lock after sleep_q.lock release to
> > >>> avoid locks nesting.
> > >
> > >>> rtw_free_xmitframe_queue(pxmitpriv, &psta->sleep_q);
> > >>> + spin_lock_bh(&psta->sleep_q.lock);
> > >>
> > >> AFAICT this needs to be above the rtw_free_xmitframe_queue() ?
> > >
> > > as I wrote on the changelog, the sleep_q.lock is already
> > > taken inside rtw_free_xmitframe_queue. If I put the
> > > sleep_q.lock above that function a soft lock occurs when
> > > I disconnect.
> > >
> > > So I put it just below rtw_free_xmitframe_queue.
> > >
> > > Things works fine this way.
> > >
> > > Please tell me if there's a best way to do it.
> >
> > Hmm I see, this may work, but the sleepq_len access
> > really should be protected by the same lock as the freeing
> > of the queue is without dropping it in between.
> >
> > That rtw_free_xmitframe_queue() takes the sleep_q.lock
> > then to me that signals that other (higher-level) functions should
> > not take sleep_q.lock at all, since this is then private to the
> > functions operating on the sleep_q.
> >
> > I've an idea how we we can possibly tackle this, but I'm not sure
> > yet I will try to make some time to look into this tomorrow or
> > the day after.
>
> I'm just going to go and revert the original change here until you all
> can sort it out :)
that's the best thing for now ;)
>
> thanks,
>
> greg k-h
thank you,
fabio