ALSA core info race condition
diff mbox

Message ID 70ff31e5f7fb35774cc0a9f3764ed144@codeaurora.org
State New
Headers show

Commit Message

b_lkasam@codeaurora.org June 27, 2017, 7:12 p.m. UTC
hi ALSA team,
there is a race condition in below API when accessing list API.

In file sound/core/info.c:

Added below patch to avoid list access of same parent node
by two threads at same time causing list_debug crash.


Please check above logic looks fine, and help comment accordingly.


Thanks
Kasam

Comments

Takashi Iwai June 28, 2017, 5:36 a.m. UTC | #1
On Tue, 27 Jun 2017 21:12:18 +0200,
b_lkasam@codeaurora.org wrote:
> 
> hi ALSA team,
> there is a race condition in below API when accessing list API.
> 
> In file sound/core/info.c:
> 
> Added below patch to avoid list access of same parent node
> by two threads at same time causing list_debug crash.
> 
> diff --git a/sound/core/info.c b/sound/core/info.c
> index b5158b5..c1fd671 100644
> --- a/sound/core/info.c
> +++ b/sound/core/info.c
> @@ -747,8 +747,11 @@ snd_info_create_entry(const char *name, struct
> snd_info_entry *parent)
>  	INIT_LIST_HEAD(&entry->children);
>  	INIT_LIST_HEAD(&entry->list);
>  	entry->parent = parent;
> -	if (parent)
> +	if (parent) {
> +		mutex_lock(&parent->access);
>  		list_add_tail(&entry->list, &parent->children);
> +		mutex_unlock(&parent->access);
> +	}
>  	return entry;
>  }
> 
> Please check above logic looks fine, and help comment accordingly.

Have you ever got the actual crash?
The function is supposed to be called only at probing, and its link
base is the card object, so it's never called concurrently or the
concurrency should be managed in the caller side.

Your "fix" looks OK, but it's likely superfluous from the actual
usage.  Still it might be safer to add a protection, though.


thanks,

Takashi
b_lkasam@codeaurora.org June 28, 2017, 5:58 a.m. UTC | #2
On 2017-06-28 11:06, Takashi Iwai wrote:
> On Tue, 27 Jun 2017 21:12:18 +0200,
> b_lkasam@codeaurora.org wrote:
>> 
>> hi ALSA team,
>> there is a race condition in below API when accessing list API.
>> 
>> In file sound/core/info.c:
>> 
>> Added below patch to avoid list access of same parent node
>> by two threads at same time causing list_debug crash.
>> 
>> diff --git a/sound/core/info.c b/sound/core/info.c
>> index b5158b5..c1fd671 100644
>> --- a/sound/core/info.c
>> +++ b/sound/core/info.c
>> @@ -747,8 +747,11 @@ snd_info_create_entry(const char *name, struct
>> snd_info_entry *parent)
>>  	INIT_LIST_HEAD(&entry->children);
>>  	INIT_LIST_HEAD(&entry->list);
>>  	entry->parent = parent;
>> -	if (parent)
>> +	if (parent) {
>> +		mutex_lock(&parent->access);
>>  		list_add_tail(&entry->list, &parent->children);
>> +		mutex_unlock(&parent->access);
>> +	}
>>  	return entry;
>>  }
>> 
>> Please check above logic looks fine, and help comment accordingly.
> 
> Have you ever got the actual crash?
> The function is supposed to be called only at probing, and its link
> base is the card object, so it's never called concurrently or the
> concurrency should be managed in the caller side.
> 
> Your "fix" looks OK, but it's likely superfluous from the actual
> usage.  Still it might be safer to add a protection, though.
> 
> 
> thanks,
> 
> Takashi


Yes Takashi, we found this crash happened on Qualcomm platform.
And as mentioned by you...it is used at bootup probe but from two 
different contexts,
Even if I try to manage at client level ...it will be only workaround. 
so actual fix
should be the patch I shared to avoid any similar issues in future.

thanks
Laxminath
Takashi Iwai June 28, 2017, 6:26 a.m. UTC | #3
On Wed, 28 Jun 2017 07:58:07 +0200,
b_lkasam@codeaurora.org wrote:
> 
> On 2017-06-28 11:06, Takashi Iwai wrote:
> > On Tue, 27 Jun 2017 21:12:18 +0200,
> > b_lkasam@codeaurora.org wrote:
> >>
> >> hi ALSA team,
> >> there is a race condition in below API when accessing list API.
> >>
> >> In file sound/core/info.c:
> >>
> >> Added below patch to avoid list access of same parent node
> >> by two threads at same time causing list_debug crash.
> >>
> >> diff --git a/sound/core/info.c b/sound/core/info.c
> >> index b5158b5..c1fd671 100644
> >> --- a/sound/core/info.c
> >> +++ b/sound/core/info.c
> >> @@ -747,8 +747,11 @@ snd_info_create_entry(const char *name, struct
> >> snd_info_entry *parent)
> >>  	INIT_LIST_HEAD(&entry->children);
> >>  	INIT_LIST_HEAD(&entry->list);
> >>  	entry->parent = parent;
> >> -	if (parent)
> >> +	if (parent) {
> >> +		mutex_lock(&parent->access);
> >>  		list_add_tail(&entry->list, &parent->children);
> >> +		mutex_unlock(&parent->access);
> >> +	}
> >>  	return entry;
> >>  }
> >>
> >> Please check above logic looks fine, and help comment accordingly.
> >
> > Have you ever got the actual crash?
> > The function is supposed to be called only at probing, and its link
> > base is the card object, so it's never called concurrently or the
> > concurrency should be managed in the caller side.
> >
> > Your "fix" looks OK, but it's likely superfluous from the actual
> > usage.  Still it might be safer to add a protection, though.
> >
> >
> > thanks,
> >
> > Takashi
> 
> 
> Yes Takashi, we found this crash happened on Qualcomm platform.
> And as mentioned by you...it is used at bootup probe but from two
> different contexts,
> Even if I try to manage at client level ...it will be only
> workaround. so actual fix
> should be the patch I shared to avoid any similar issues in future.

Hrm, it already smells suspicious.  The situation where the proc file
creation on the same root is called concurrently implies the racy
registration, not only about the proc...

But it makes no sense to discuss without the actual code, and the
suggested change looks OK, as said.  So please submit a proper patch
in the usual manner.


thanks,

Takashi

Patch
diff mbox

diff --git a/sound/core/info.c b/sound/core/info.c
index b5158b5..c1fd671 100644
--- a/sound/core/info.c
+++ b/sound/core/info.c
@@ -747,8 +747,11 @@  snd_info_create_entry(const char *name, struct 
snd_info_entry *parent)
  	INIT_LIST_HEAD(&entry->children);
  	INIT_LIST_HEAD(&entry->list);
  	entry->parent = parent;
-	if (parent)
+	if (parent) {
+		mutex_lock(&parent->access);
  		list_add_tail(&entry->list, &parent->children);
+		mutex_unlock(&parent->access);
+	}
  	return entry;
  }