ipv4: initialize ra_mutex in inet_init_net()
Subsystems (2)
PATCH 1/1 ipv4: initialize ra_mutex in inet_init_net()
ra_mutex is a IPv4 specific mutex, it is inside struct netns_ipv4, but its initialization is in the generic netns code, setup_net().
Move it to IPv4 specific net init code, inet_init_net().
--- net/core/net_namespace.c | 1 - net/ipv4/af_inet.c | 2 ++ 2 files changed, 2 insertions(+), 1 deletion(-)
net/core/net_namespace.c +0 -1
@@ -308,7 +308,6 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns) net->user_ns = user_ns;idr_init (&net->netns_ids);spin_lock_init (&net->nsid_lock); -mutex_init (&net->ipv4.ra_mutex);list_for_each_entry (ops, &pernet_list, list) { error =ops_init (ops, net);
net/ipv4/af_inet.c +2 -0
@@ -1817,6 +1817,8 @@ static __net_init int inet_init_net(struct net *net) net->ipv4.sysctl_igmp_llm_reports =1 ; net->ipv4.sysctl_igmp_qrv =2 ; +mutex_init (&net->ipv4.ra_mutex); +return 0 ; }
Signature · 1 line
2.14.4
Quoted 30 lines · Cong Wang
ra_mutexis a IPv4 specific mutex, it is inside structnetns_ipv4, but its initialization is in the generic netns code,setup_net().Move it to IPv4 specific net init code,
inet_init_net().Fixes d9ff3049739e ("net: Replace ip_ra_lock with per-net mutex")Cc Kirill Tkhai <[email protected]>Signed-off-by Cong Wang <[email protected]>--- net/core/net_namespace.c | 1 - net/ipv4/af_inet.c | 2 ++ 2 files changed, 2 insertions(+), 1 deletion(-)net/core/net_namespace.c +0
-1@@ -308,7 +308,6 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns) net->user_ns = user_ns;idr_init (&net->netns_ids);spin_lock_init (&net->nsid_lock); -mutex_init (&net->ipv4.ra_mutex);list_for_each_entry (ops, &pernet_list, list) { error =ops_init (ops, net);net/ipv4/af_inet.c +0
-0
net/ipv4/af_inet.c static __net_init int inet_init_net(struct net *net) · line 1821
Quoted 5 lines
net/ipv4/af_inet.c +2
-0@@ -1817,6 +1817,8 @@ static __net_init int inet_init_net(struct net *net) net->ipv4.sysctl_igmp_llm_reports =1 ; net->ipv4.sysctl_igmp_qrv =2 ; +mutex_init (&net->ipv4.ra_mutex); +
In inet_init() the order of registration is:
ip_mr_init(); init_inet_pernet_ops();
This means, ipmr_net_ops pernet operations are before af_inet_ops in pernet_list. So, there is a theoretical probability, sometimes in the future, we will have a problem during a fail of net initialization.
Say,
setup_net():
ipmr_net_ops->init() returns 0 xxx->init() returns error and then we do: ipmr_net_ops->exit(),
which could touch ra_mutex (theoretically).
Your patch is OK, but since you do this, we may also swap the order of registration of ipmr_net_ops and af_inet_ops better too.
Kirill
Quoted 11 lines · Kirill Tkhai
In
inet_init()the order of registration is:ip_mr_init(); init_inet_pernet_ops();This means,
ipmr_net_opspernet operations are beforeaf_inet_opsinpernet_list. So, there is a theoretical probability, sometimes in the future, we will have a problem during a fail of net initialization.Say,
Quoted 7 lines
setup_net():ipmr_net_ops->init() returns 0 xxx->init() returns error and then we do: ipmr_net_ops->exit(),which could touch
ra_mutex(theoretically).
How could ra_mutex be touched in this scenario?
ra_mutex is only used in ip_ra_control() which is called only by {get,set}sockopt(). I don't see anything related to netns exit() path here.
Quoted 18 lines · Cong Wang
Quoted 17 lines · Kirill Tkhai
In
inet_init()the order of registration is:ip_mr_init(); init_inet_pernet_ops();This means,
ipmr_net_opspernet operations are beforeaf_inet_opsinpernet_list. So, there is a theoretical probability, sometimes in the future, we will have a problem during a fail of net initialization.Say,
setup_net():ipmr_net_ops->init() returns 0 xxx->init() returns error and then we do: ipmr_net_ops->exit(),
Quoted 7 lines
Quoted 1 line
which could touch
ra_mutex(theoretically).How could
ra_mutexbe touched in this scenario?
ra_mutexis only used inip_ra_control()which is called only by {get,set}sockopt(). I don't see anything related to netnsexit()path here.
Currently, it is not touched. But it's an ordinary practice, someone closes sockets in pernet ->exit methods. For example, we close percpu icmp sockets in icmp_sk_exit(), which are also of RAW type, and there is also called ip_ra_control() for them. Yes, they differ by their protocol; icmp sockets are of IPPROTO_ICMP protocol, while ip_ra_control() acts on IPPROTO_RAW sockets, but it's not good anyway. This does not look reliable for the future. In case of someone changes something here, we may do not notice this for the long time, while some users will meet bugs on their places.
Problems on error paths is not easy to detect on testing, while user may meet them. We had issue of same type with uninitialized xfrm_policy_lock. It was introduced in 2013, while the problem was found only in 2017:
introduced by 283bc9f35bbb fixed by c282222a45cb
(Last week I met it on RH7 kernel, which still has no a fix.
But this talk is not about distribution kernels, just about the way).
I just want to say if someone makes some creativity on top of this code, it will be to more friendly from us to him/her to not force this person to think about such not obvious details, but just to implement nice architecture right now.
Thanks,
Kirill
Quoted 28 lines · Kirill Tkhai
Quoted 18 lines · Cong Wang
Quoted 17 lines · Kirill Tkhai
In
inet_init()the order of registration is:ip_mr_init(); init_inet_pernet_ops();This means,
ipmr_net_opspernet operations are beforeaf_inet_opsinpernet_list. So, there is a theoretical probability, sometimes in the future, we will have a problem during a fail of net initialization.Say,
setup_net():ipmr_net_ops->init() returns 0 xxx->init() returns error and then we do: ipmr_net_ops->exit(),Quoted 7 lines
Quoted 1 line
which could touch
ra_mutex(theoretically).How could
ra_mutexbe touched in this scenario?
ra_mutexis only used inip_ra_control()which is called only by {get,set}sockopt(). I don't see anything related to netnsexit()path here.
Quoted 10 lines
Currently, it is not touched. But it's an ordinary practice, someone closes sockets in pernet ->exit methods. For example, we close percpu icmp sockets in
icmp_sk_exit(), which are also of RAW type, and there is also calledip_ra_control()for them. Yes, they differ by their protocol; icmp sockets are ofIPPROTO_ICMPprotocol, whileip_ra_control()acts onIPPROTO_RAWsockets, but it's not good anyway. This does not look reliable for the future. In case of someone changes something here, we may do not notice this for the long time, while some users will meet bugs on their places.
First of all, we only consider current code base. Even if you really planned to changed this in the future, it would be still your responsibility to take care of it. Why? Simple, FIFO. My patch comes ahead of any future changes here, obviously.
Secondly, it is certainly not hard to notice. I am pretty sure you would get a warning/crash if this bug triggered.
Quoted 5 lines
Problems on error paths is not easy to detect on testing, while user may meet them. We had issue of same type with uninitialized
xfrm_policy_lock. It was introduced in 2013, while the problem was found only in 2017:
Been there, done that, I've fixed multiple IPv6 init code error path. This is not where we disagree, by the way.
Quoted 4 lines
introduced by
283bc9f35bbb
fixed byc282222a45cb
Quoted 8 lines
(Last week I met it on RH7 kernel, which still has no a fix.
But this talk is not about distribution kernels, just about the way).I just want to say if someone makes some creativity on top of this code, it will be to more friendly from us to him/her to not force this person to think about such not obvious details, but just to implement nice architecture right now.
You keep saying nice architecture, how did you architect ipv4.ra_mutex into net/core/net_namespace.c? It is apparently not nice.
Good luck with your future.
I am tired, let's just drop it. I have no interest to fix your mess.
Thanks!
Quoted 73 lines · Cong Wang
Quoted 28 lines · Kirill Tkhai
Quoted 18 lines · Cong Wang
Quoted 17 lines · Kirill Tkhai
In
inet_init()the order of registration is:ip_mr_init(); init_inet_pernet_ops();This means,
ipmr_net_opspernet operations are beforeaf_inet_opsinpernet_list. So, there is a theoretical probability, sometimes in the future, we will have a problem during a fail of net initialization.Say,
setup_net():ipmr_net_ops->init() returns 0 xxx->init() returns error and then we do: ipmr_net_ops->exit(),Quoted 7 lines
Quoted 1 line
which could touch
ra_mutex(theoretically).How could
ra_mutexbe touched in this scenario?
ra_mutexis only used inip_ra_control()which is called only by {get,set}sockopt(). I don't see anything related to netnsexit()path here.Quoted 10 lines
Currently, it is not touched. But it's an ordinary practice, someone closes sockets in pernet ->exit methods. For example, we close percpu icmp sockets in
icmp_sk_exit(), which are also of RAW type, and there is also calledip_ra_control()for them. Yes, they differ by their protocol; icmp sockets are ofIPPROTO_ICMPprotocol, whileip_ra_control()acts onIPPROTO_RAWsockets, but it's not good anyway. This does not look reliable for the future. In case of someone changes something here, we may do not notice this for the long time, while some users will meet bugs on their places.First of all, we only consider current code base. Even if you really planned to changed this in the future, it would be still your responsibility to take care of it. Why? Simple, FIFO. My patch comes ahead of any future changes here, obviously.
Secondly, it is certainly not hard to notice. I am pretty sure you would get a warning/crash if this bug triggered.
Quoted 5 lines
Problems on error paths is not easy to detect on testing, while user may meet them. We had issue of same type with uninitialized
xfrm_policy_lock. It was introduced in 2013, while the problem was found only in 2017:Been there, done that, I've fixed multiple IPv6 init code error path. This is not where we disagree, by the way.
Quoted 4 lines
introduced by
283bc9f35bbb
fixed byc282222a45cbQuoted 8 lines
(Last week I met it on RH7 kernel, which still has no a fix.
But this talk is not about distribution kernels, just about the way).I just want to say if someone makes some creativity on top of this code, it will be to more friendly from us to him/her to not force this person to think about such not obvious details, but just to implement nice architecture right now.
Quoted 7 lines
You keep saying nice architecture, how did you architect ipv4.ra_mutex into
net/core/net_namespace.c? It is apparently not nice.Good luck with your future.
I am tired, let's just drop it. I have no interest to fix your mess.
You added me to CC, so you probably want to know my opinion about this. Since it's not a real problem fix, but just a refactoring, I say you my opinion, how this refactoring may be made better. If you don't want to know my opinion, you may consider not to CC me.
Just this, not a reason to take offense.
Thanks,
Kirill
Quoted 6 lines · Kirill Tkhai
On 20.09.2018 0:28, Cong Wang wrote:
You added me to CC, so you probably want to know my opinion about this. Since it's not a real problem fix, but just a refactoring, I say you my opinion, how this refactoring may be made better. If you don't want to know my opinion, you may consider not to CC me.
Sure, with respects to your opinions, I prefer to drop it happily.
Thanks.
David Miller
maintainerFrom: Cong Wang <[email protected]>
Date: Fri, 14 Sep 2018 13:32:42 -0700
Quoted 8 lines
ra_mutexis a IPv4 specific mutex, it is inside structnetns_ipv4, but its initialization is in the generic netns code,setup_net().Move it to IPv4 specific net init code,
inet_init_net().Fixes d9ff3049739e ("net: Replace ip_ra_lock with per-net mutex")Cc Kirill Tkhai <[email protected]>Signed-off-by Cong Wang <[email protected]>
Please take into consideration Kirill's feedback.
Thank you.