mmotm 2010-04-28 - RCU whinges

David Miller wrote:
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Mon, 03 May 2010 07:43:56 +0200
> 
>> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
>>
>>> Oops scratch that, I'll resend a correct version.
>>>
>>>
>> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
>> thought a different mutex was in use in one of the functions.
> 
> Ok, Patrick please review, thanks.

Actually we don't need the rcu_dereference() calls at all since
registration and unregistration are protected by the mutexes.

I queued this patch in nf-next, the only reason why I haven't
submitted it yet is that I was unable to get git to cleanly export
only the proper set of patches meant for -next due to a few merges,
it insists on including 5 patches already merged upstream. If you
don't mind ignoring the first 5 patches in the series, I'll send a
pull request tonight.
commit ed86308f6179d8fa6151c2d0f652aad0091548e2
Author: Patrick McHardy <kaber@trash.net>
Date:   Fri Apr 9 16:42:15 2010 +0200

    netfilter: remove invalid rcu_dereference() calls
    
    The CONFIG_PROVE_RCU option discovered a few invalid uses of
    rcu_dereference() in netfilter. In all these cases, the code code
    intends to check whether a pointer is already assigned when
    performing registration or whether the assigned pointer matches
    when performing unregistration. The entire registration/
    unregistration is protected by a mutex, so we don't need the
    rcu_dereference() calls.
    
    Reported-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
    Tested-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
    Signed-off-by: Patrick McHardy <kaber@trash.net>

Le lundi 10 mai 2010 à 17:40 +0200, Patrick McHardy a écrit :
> David Miller wrote:
> > From: Eric Dumazet <eric.dumazet@gmail.com>
> > Date: Mon, 03 May 2010 07:43:56 +0200
> > 
> >> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
> >>
> >>> Oops scratch that, I'll resend a correct version.
> >>>
> >>>
> >> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
> >> thought a different mutex was in use in one of the functions.
> > 
> > Ok, Patrick please review, thanks.
> 
> Actually we don't need the rcu_dereference() calls at all since
> registration and unregistration are protected by the mutexes.
> 
> I queued this patch in nf-next, the only reason why I haven't
> submitted it yet is that I was unable to get git to cleanly export
> only the proper set of patches meant for -next due to a few merges,
> it insists on including 5 patches already merged upstream. If you
> don't mind ignoring the first 5 patches in the series, I'll send a
> pull request tonight.
> 


This will clash with upcoming RCU patches, where rcu protected pointer
cannot be directly accessed without lockdep splats.

We will need one day or another a rcu_...(nf_conntrack_event_cb)

> 
> pièce jointe document texte brut (x)
> commit ed86308f6179d8fa6151c2d0f652aad0091548e2
> Author: Patrick McHardy <kaber@trash.net>
> Date:   Fri Apr 9 16:42:15 2010 +0200
> 
>     netfilter: remove invalid rcu_dereference() calls
>     
>     The CONFIG_PROVE_RCU option discovered a few invalid uses of
>     rcu_dereference() in netfilter. In all these cases, the code code
>     intends to check whether a pointer is already assigned when
>     performing registration or whether the assigned pointer matches
>     when performing unregistration. The entire registration/
>     unregistration is protected by a mutex, so we don't need the
>     rcu_dereference() calls.
>     
>     Reported-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
>     Tested-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
>     Signed-off-by: Patrick McHardy <kaber@trash.net>
> 
> diff --git a/net/netfilter/nf_conntrack_ecache.c b/net/netfilter/nf_conntrack_ecache.c
> index d5a9bcd..849614a 100644
> --- a/net/netfilter/nf_conntrack_ecache.c
> +++ b/net/netfilter/nf_conntrack_ecache.c
> @@ -81,11 +81,9 @@ EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events);
>  int nf_conntrack_register_notifier(struct nf_ct_event_notifier *new)
>  {
>  	int ret = 0;
> -	struct nf_ct_event_notifier *notify;
>  
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_conntrack_event_cb);
> -	if (notify != NULL) {
> +	if (nf_conntrack_event_cb != NULL) {
>  		ret = -EBUSY;
>  		goto out_unlock;
>  	}
> @@ -101,11 +99,8 @@ EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier);
>  
>  void nf_conntrack_unregister_notifier(struct nf_ct_event_notifier *new)
>  {
> -	struct nf_ct_event_notifier *notify;
> -
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_conntrack_event_cb);
> -	BUG_ON(notify != new);
> +	BUG_ON(nf_conntrack_event_cb != new);
>  	rcu_assign_pointer(nf_conntrack_event_cb, NULL);
>  	mutex_unlock(&nf_ct_ecache_mutex);
>  }
> @@ -114,11 +109,9 @@ EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier);
>  int nf_ct_expect_register_notifier(struct nf_exp_event_notifier *new)
>  {
>  	int ret = 0;
> -	struct nf_exp_event_notifier *notify;
>  
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_expect_event_cb);
> -	if (notify != NULL) {
> +	if (nf_expect_event_cb != NULL) {
>  		ret = -EBUSY;
>  		goto out_unlock;
>  	}
> @@ -134,11 +127,8 @@ EXPORT_SYMBOL_GPL(nf_ct_expect_register_notifier);
>  
>  void nf_ct_expect_unregister_notifier(struct nf_exp_event_notifier *new)
>  {
> -	struct nf_exp_event_notifier *notify;
> -
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_expect_event_cb);
> -	BUG_ON(notify != new);
> +	BUG_ON(nf_expect_event_cb != new);
>  	rcu_assign_pointer(nf_expect_event_cb, NULL);
>  	mutex_unlock(&nf_ct_ecache_mutex);
>  }
> diff --git a/net/netfilter/nf_log.c b/net/netfilter/nf_log.c
> index 015725a..908f599 100644
> --- a/net/netfilter/nf_log.c
> +++ b/net/netfilter/nf_log.c
> @@ -35,7 +35,6 @@ static struct nf_logger *__find_logger(int pf, const char *str_logger)
>  /* return EEXIST if the same logger is registred, 0 on success. */
>  int nf_log_register(u_int8_t pf, struct nf_logger *logger)
>  {
> -	const struct nf_logger *llog;
>  	int i;
>  
>  	if (pf >= ARRAY_SIZE(nf_loggers))
> @@ -52,8 +51,7 @@ int nf_log_register(u_int8_t pf, struct nf_logger *logger)
>  	} else {
>  		/* register at end of list to honor first register win */
>  		list_add_tail(&logger->list[pf], &nf_loggers_l[pf]);
> -		llog = rcu_dereference(nf_loggers[pf]);
> -		if (llog == NULL)
> +		if (nf_loggers[pf] == NULL)
>  			rcu_assign_pointer(nf_loggers[pf], logger);
>  	}
>  
> @@ -65,13 +63,11 @@ EXPORT_SYMBOL(nf_log_register);
>  
>  void nf_log_unregister(struct nf_logger *logger)
>  {
> -	const struct nf_logger *c_logger;
>  	int i;
>  
>  	mutex_lock(&nf_log_mutex);
>  	for (i = 0; i < ARRAY_SIZE(nf_loggers); i++) {
> -		c_logger = rcu_dereference(nf_loggers[i]);
> -		if (c_logger == logger)
> +		if (nf_loggers[i] == logger)
>  			rcu_assign_pointer(nf_loggers[i], NULL);
>  		list_del(&logger->list[i]);
>  	}


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Le lundi 10 mai 2010 à 17:56 +0200, Eric Dumazet a écrit :
> Le lundi 10 mai 2010 à 17:40 +0200, Patrick McHardy a écrit :
> > David Miller wrote:
> > > From: Eric Dumazet <eric.dumazet@gmail.com>
> > > Date: Mon, 03 May 2010 07:43:56 +0200
> > > 
> > >> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
> > >>
> > >>> Oops scratch that, I'll resend a correct version.
> > >>>
> > >>>
> > >> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
> > >> thought a different mutex was in use in one of the functions.
> > > 
> > > Ok, Patrick please review, thanks.
> > 
> > Actually we don't need the rcu_dereference() calls at all since
> > registration and unregistration are protected by the mutexes.
> > 
> > I queued this patch in nf-next, the only reason why I haven't
> > submitted it yet is that I was unable to get git to cleanly export
> > only the proper set of patches meant for -next due to a few merges,
> > it insists on including 5 patches already merged upstream. If you
> > don't mind ignoring the first 5 patches in the series, I'll send a
> > pull request tonight.
> > 
> 
> 
> This will clash with upcoming RCU patches, where rcu protected pointer
> cannot be directly accessed without lockdep splats.
> 

Sorry, I meant sparse here, not lockdep.

> We will need one day or another a rcu_...(nf_conntrack_event_cb)


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Mon, May 10, 2010 at 05:40:58PM +0200, Patrick McHardy wrote:
> David Miller wrote:
> > From: Eric Dumazet <eric.dumazet@gmail.com>
> > Date: Mon, 03 May 2010 07:43:56 +0200
> > 
> >> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
> >>
> >>> Oops scratch that, I'll resend a correct version.
> >>>
> >>>
> >> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
> >> thought a different mutex was in use in one of the functions.
> > 
> > Ok, Patrick please review, thanks.
> 
> Actually we don't need the rcu_dereference() calls at all since
> registration and unregistration are protected by the mutexes.

The best approach in that case is rcu_dereference_protected() listing
the lock that must be held.  Of course, your code, so your choice.

						Thanx, Paul

> I queued this patch in nf-next, the only reason why I haven't
> submitted it yet is that I was unable to get git to cleanly export
> only the proper set of patches meant for -next due to a few merges,
> it insists on including 5 patches already merged upstream. If you
> don't mind ignoring the first 5 patches in the series, I'll send a
> pull request tonight.
> 
> 

> commit ed86308f6179d8fa6151c2d0f652aad0091548e2
> Author: Patrick McHardy <kaber@trash.net>
> Date:   Fri Apr 9 16:42:15 2010 +0200
> 
>     netfilter: remove invalid rcu_dereference() calls
>     
>     The CONFIG_PROVE_RCU option discovered a few invalid uses of
>     rcu_dereference() in netfilter. In all these cases, the code code
>     intends to check whether a pointer is already assigned when
>     performing registration or whether the assigned pointer matches
>     when performing unregistration. The entire registration/
>     unregistration is protected by a mutex, so we don't need the
>     rcu_dereference() calls.
>     
>     Reported-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
>     Tested-by: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
>     Signed-off-by: Patrick McHardy <kaber@trash.net>
> 
> diff --git a/net/netfilter/nf_conntrack_ecache.c b/net/netfilter/nf_conntrack_ecache.c
> index d5a9bcd..849614a 100644
> --- a/net/netfilter/nf_conntrack_ecache.c
> +++ b/net/netfilter/nf_conntrack_ecache.c
> @@ -81,11 +81,9 @@ EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events);
>  int nf_conntrack_register_notifier(struct nf_ct_event_notifier *new)
>  {
>  	int ret = 0;
> -	struct nf_ct_event_notifier *notify;
> 
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_conntrack_event_cb);
> -	if (notify != NULL) {
> +	if (nf_conntrack_event_cb != NULL) {
>  		ret = -EBUSY;
>  		goto out_unlock;
>  	}
> @@ -101,11 +99,8 @@ EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier);
> 
>  void nf_conntrack_unregister_notifier(struct nf_ct_event_notifier *new)
>  {
> -	struct nf_ct_event_notifier *notify;
> -
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_conntrack_event_cb);
> -	BUG_ON(notify != new);
> +	BUG_ON(nf_conntrack_event_cb != new);
>  	rcu_assign_pointer(nf_conntrack_event_cb, NULL);
>  	mutex_unlock(&nf_ct_ecache_mutex);
>  }
> @@ -114,11 +109,9 @@ EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier);
>  int nf_ct_expect_register_notifier(struct nf_exp_event_notifier *new)
>  {
>  	int ret = 0;
> -	struct nf_exp_event_notifier *notify;
> 
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_expect_event_cb);
> -	if (notify != NULL) {
> +	if (nf_expect_event_cb != NULL) {
>  		ret = -EBUSY;
>  		goto out_unlock;
>  	}
> @@ -134,11 +127,8 @@ EXPORT_SYMBOL_GPL(nf_ct_expect_register_notifier);
> 
>  void nf_ct_expect_unregister_notifier(struct nf_exp_event_notifier *new)
>  {
> -	struct nf_exp_event_notifier *notify;
> -
>  	mutex_lock(&nf_ct_ecache_mutex);
> -	notify = rcu_dereference(nf_expect_event_cb);
> -	BUG_ON(notify != new);
> +	BUG_ON(nf_expect_event_cb != new);
>  	rcu_assign_pointer(nf_expect_event_cb, NULL);
>  	mutex_unlock(&nf_ct_ecache_mutex);
>  }
> diff --git a/net/netfilter/nf_log.c b/net/netfilter/nf_log.c
> index 015725a..908f599 100644
> --- a/net/netfilter/nf_log.c
> +++ b/net/netfilter/nf_log.c
> @@ -35,7 +35,6 @@ static struct nf_logger *__find_logger(int pf, const char *str_logger)
>  /* return EEXIST if the same logger is registred, 0 on success. */
>  int nf_log_register(u_int8_t pf, struct nf_logger *logger)
>  {
> -	const struct nf_logger *llog;
>  	int i;
> 
>  	if (pf >= ARRAY_SIZE(nf_loggers))
> @@ -52,8 +51,7 @@ int nf_log_register(u_int8_t pf, struct nf_logger *logger)
>  	} else {
>  		/* register at end of list to honor first register win */
>  		list_add_tail(&logger->list[pf], &nf_loggers_l[pf]);
> -		llog = rcu_dereference(nf_loggers[pf]);
> -		if (llog == NULL)
> +		if (nf_loggers[pf] == NULL)
>  			rcu_assign_pointer(nf_loggers[pf], logger);
>  	}
> 
> @@ -65,13 +63,11 @@ EXPORT_SYMBOL(nf_log_register);
> 
>  void nf_log_unregister(struct nf_logger *logger)
>  {
> -	const struct nf_logger *c_logger;
>  	int i;
> 
>  	mutex_lock(&nf_log_mutex);
>  	for (i = 0; i < ARRAY_SIZE(nf_loggers); i++) {
> -		c_logger = rcu_dereference(nf_loggers[i]);
> -		if (c_logger == logger)
> +		if (nf_loggers[i] == logger)
>  			rcu_assign_pointer(nf_loggers[i], NULL);
>  		list_del(&logger->list[i]);
>  	}

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Eric Dumazet wrote:
> Le lundi 10 mai 2010 à 17:40 +0200, Patrick McHardy a écrit :
>> David Miller wrote:
>>> From: Eric Dumazet <eric.dumazet@gmail.com>
>>> Date: Mon, 03 May 2010 07:43:56 +0200
>>>
>>>> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
>>>>
>>>>> Oops scratch that, I'll resend a correct version.
>>>>>
>>>>>
>>>> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
>>>> thought a different mutex was in use in one of the functions.
>>> Ok, Patrick please review, thanks.
>> Actually we don't need the rcu_dereference() calls at all since
>> registration and unregistration are protected by the mutexes.
>>
>> I queued this patch in nf-next, the only reason why I haven't
>> submitted it yet is that I was unable to get git to cleanly export
>> only the proper set of patches meant for -next due to a few merges,
>> it insists on including 5 patches already merged upstream. If you
>> don't mind ignoring the first 5 patches in the series, I'll send a
>> pull request tonight.
>>
> 
> This will clash with upcoming RCU patches, where rcu protected pointer
> cannot be directly accessed without lockdep splats.
> 
> We will need one day or another a rcu_...(nf_conntrack_event_cb)

Thanks for the information, I didn't realize that when looking at
those patches. So I guess the correct fix once those patches are
merged would be to use rcu_assign_protected() and rcu_access_pointer().
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patrick McHardy wrote:
> Eric Dumazet wrote:
>> Le lundi 10 mai 2010 à 17:40 +0200, Patrick McHardy a écrit :
>>> David Miller wrote:
>>>> From: Eric Dumazet <eric.dumazet@gmail.com>
>>>> Date: Mon, 03 May 2010 07:43:56 +0200
>>>>
>>>>> Le lundi 03 mai 2010 à 07:41 +0200, Eric Dumazet a écrit :
>>>>>
>>>>>> Oops scratch that, I'll resend a correct version.
>>>>>>
>>>>>>
>>>>> Sorry, patch _is_ fine, I had one brain collapse when re-reading it, I
>>>>> thought a different mutex was in use in one of the functions.
>>>> Ok, Patrick please review, thanks.
>>> Actually we don't need the rcu_dereference() calls at all since
>>> registration and unregistration are protected by the mutexes.
>>>
>>> I queued this patch in nf-next, the only reason why I haven't
>>> submitted it yet is that I was unable to get git to cleanly export
>>> only the proper set of patches meant for -next due to a few merges,
>>> it insists on including 5 patches already merged upstream. If you
>>> don't mind ignoring the first 5 patches in the series, I'll send a
>>> pull request tonight.
>>>
>> This will clash with upcoming RCU patches, where rcu protected pointer
>> cannot be directly accessed without lockdep splats.
>>
>> We will need one day or another a rcu_...(nf_conntrack_event_cb)
> 
> Thanks for the information, I didn't realize that when looking at
> those patches. So I guess the correct fix once those patches are
> merged would be to use rcu_assign_protected() and rcu_access_pointer().

Ah, and that's what you did. Sorry for the confusion :)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

From: Patrick McHardy <kaber@trash.net>
Date: Mon, 10 May 2010 17:40:58 +0200

> I queued this patch in nf-next, the only reason why I haven't
> submitted it yet is that I was unable to get git to cleanly export
> only the proper set of patches meant for -next due to a few merges,
> it insists on including 5 patches already merged upstream. If you
> don't mind ignoring the first 5 patches in the series, I'll send a
> pull request tonight.

Something like "git format-patch origin" doesn't avoid those upstream
commits?  Weird...

Another trick is to specify a commit range using triple-dot "..."
notation, such as "origin...master"
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

David Miller wrote:
> From: Patrick McHardy <kaber@trash.net>
> Date: Mon, 10 May 2010 17:40:58 +0200
> 
>> I queued this patch in nf-next, the only reason why I haven't
>> submitted it yet is that I was unable to get git to cleanly export
>> only the proper set of patches meant for -next due to a few merges,
>> it insists on including 5 patches already merged upstream. If you
>> don't mind ignoring the first 5 patches in the series, I'll send a
>> pull request tonight.
> 
> Something like "git format-patch origin" doesn't avoid those upstream
> commits?  Weird...

Yeah, it seems to have something to do with me merging the nf-2.6.git
tree a few weeks ago since it had patches queued that were too late
for 2.6.34. Even the --ignore-if-in-upstream option doesn't help.

> Another trick is to specify a commit range using triple-dot "..."
> notation, such as "origin...master"

Thanks, I'll give it another try, the alternative is manually
renumbering the entire patchset.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html