* [PATCH v3 ipsec-next] xfrm: remove VLA usage in __xfrm6_sort()
From: Kees Cook @ 2018-04-24 23:46 UTC (permalink / raw)
To: Stefano Brivio
Cc: Andreas Christoforou, kernel-hardening, Steffen Klassert,
Herbert Xu, David S. Miller, Alexey Kuznetsov, Hideaki YOSHIFUJI,
netdev, linux-kernel
In the quest to remove all stack VLA usage removed from the kernel[1],
just use XFRM_MAX_DEPTH as already done for the "class" array. In one
case, it'll do this loop up to 5, the other caller up to 6.
[1] https://lkml.org/lkml/2018/3/7/621
Co-developed-by: Andreas Christoforou <andreaschristofo@gmail.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
---
v3:
- adjust Subject and commit log (Steffen)
- use "= { }" instead of memset() (Stefano)
- reorder variables (Stefano)
v2:
- use XFRM_MAX_DEPTH for "count" array (Steffen and Mathias).
---
net/ipv6/xfrm6_state.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/net/ipv6/xfrm6_state.c b/net/ipv6/xfrm6_state.c
index 16f434791763..eeb44b64ae7f 100644
--- a/net/ipv6/xfrm6_state.c
+++ b/net/ipv6/xfrm6_state.c
@@ -60,9 +60,9 @@ xfrm6_init_temprop(struct xfrm_state *x, const struct xfrm_tmpl *tmpl,
static int
__xfrm6_sort(void **dst, void **src, int n, int (*cmp)(void *p), int maxclass)
{
- int i;
+ int count[XFRM_MAX_DEPTH] = { };
int class[XFRM_MAX_DEPTH];
- int count[maxclass];
+ int i;
memset(count, 0, sizeof(count));
--
2.7.4
--
Kees Cook
Pixel Security
^ permalink raw reply related
* [bpf-next PATCH] bpf: reduce runtime of test_sockmap tests
From: John Fastabend @ 2018-04-24 23:28 UTC (permalink / raw)
To: ast, daniel; +Cc: netdev
When test_sockmap was running outside of selftests and was not being
run by build bots it was reasonable to spend significant amount of
time running various tests. The number of tests is high because many
different I/O iterators are run.
However, now that test_sockmap is part of selftests rather than
iterate through all I/O sides only test a minimal set of min/max
values along with a few "normal" I/O ops. Also remove the long
running tests. They can be run from other test frameworks on a regular
cadence.
This significanly reduces runtime of test_sockmap.
Before:
$ time sudo ./test_sockmap > /dev/null
real 4m47.521s
user 0m0.370s
sys 0m3.131s
After:
$ time sudo ./test_sockmap > /dev/null
real 0m0.514s
user 0m0.104s
sys 0m0.430s
The CLI is still available for users that want to test the long
running tests that do the larger send/recv tests.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
---
tools/testing/selftests/bpf/test_sockmap.c | 33 ++++++++++++++--------------
1 file changed, 16 insertions(+), 17 deletions(-)
diff --git a/tools/testing/selftests/bpf/test_sockmap.c b/tools/testing/selftests/bpf/test_sockmap.c
index 6d63a1c..29c022d 100644
--- a/tools/testing/selftests/bpf/test_sockmap.c
+++ b/tools/testing/selftests/bpf/test_sockmap.c
@@ -344,8 +344,8 @@ static int msg_loop(int fd, int iov_count, int iov_length, int cnt,
if (err < 0)
perror("recv start time: ");
while (s->bytes_recvd < total_bytes) {
- timeout.tv_sec = 1;
- timeout.tv_usec = 0;
+ timeout.tv_sec = 0;
+ timeout.tv_usec = 10;
/* FD sets */
FD_ZERO(&w);
@@ -903,12 +903,10 @@ static int test_exec(int cgrp, struct sockmap_options *opt)
{
int err = __test_exec(cgrp, SENDMSG, opt);
- sched_yield();
if (err)
goto out;
err = __test_exec(cgrp, SENDPAGE, opt);
- sched_yield();
out:
return err;
}
@@ -928,19 +926,18 @@ static int test_loop(int cgrp)
opt.iov_length = 0;
opt.rate = 0;
- for (r = 1; r < 100; r += 33) {
- for (i = 1; i < 100; i += 33) {
- for (l = 1; l < 100; l += 33) {
- opt.rate = r;
- opt.iov_count = i;
- opt.iov_length = l;
- err = test_exec(cgrp, &opt);
- if (err)
- goto out;
- }
+ r = 1;
+ for (i = 1; i < 100; i += 33) {
+ for (l = 1; l < 100; l += 33) {
+ opt.rate = r;
+ opt.iov_count = i;
+ opt.iov_length = l;
+ err = test_exec(cgrp, &opt);
+ if (err)
+ goto out;
}
}
-
+ sched_yield();
out:
return err;
}
@@ -1031,6 +1028,7 @@ static int test_send(struct sockmap_options *opt, int cgrp)
if (err)
goto out;
out:
+ sched_yield();
return err;
}
@@ -1168,7 +1166,7 @@ static int test_start_end(int cgrp)
opt.iov_length = 100;
txmsg_cork = 1600;
- for (i = 99; i <= 1600; i += 100) {
+ for (i = 99; i <= 1600; i += 500) {
txmsg_start = 0;
txmsg_end = i;
err = test_exec(cgrp, &opt);
@@ -1177,7 +1175,7 @@ static int test_start_end(int cgrp)
}
/* Test start/end with cork but pull data in middle */
- for (i = 199; i <= 1600; i += 100) {
+ for (i = 199; i <= 1600; i += 500) {
txmsg_start = 100;
txmsg_end = i;
err = test_exec(cgrp, &opt);
@@ -1221,6 +1219,7 @@ static int test_start_end(int cgrp)
out:
txmsg_start = 0;
txmsg_end = 0;
+ sched_yield();
return err;
}
^ permalink raw reply related
* Re: [PATCH] [PATCH bpf-next] samples/bpf/bpf_load.c: remove redundant ret, assignment in bpf_load_program()
From: Daniel Borkmann @ 2018-04-24 23:15 UTC (permalink / raw)
To: shhuiw, ast, netdev
In-Reply-To: <a953464a-385a-38f3-df84-d9dac58b99b4@foxmail.com>
On 04/24/2018 10:18 AM, shhuiw wrote:
>
> 2 redundant ret assignments removded:
> * 'ret = 1' before the logic 'if (data_maps)', and if any errors jump to
> label 'done'. No 'ret = 1' needed before the error jump.
> * After the '/* load programs */' part, if everything goes well, then
> the BPF code will be loaded and 'ret' set to 0 by load_and_attach().
> If something goes wrong, 'ret' set to none-O, the redundant 'ret = 0'
> after the for clause will make the error skipped.
> For example, if some BPF code cannot provide supported program types
> in ELF SEC("unknown"), the for clause will not call load_and_attach()
> to load the BPF code. 1 should be returned to callees instead of 0.
>
> Signed-off-by: Wang Sheng-Hui <shhuiw@foxmail.com>
Your patch is corrupted, please use something like git-send-email(1).
Thanks,
Daniel
^ permalink raw reply
* Re: [PATCH net-next 1/2 v2] netns: restrict uevents
From: Eric W. Biederman @ 2018-04-24 23:00 UTC (permalink / raw)
To: Christian Brauner
Cc: Christian Brauner, David Miller, netdev, linux-kernel, avagin,
ktkhai, serge, gregkh
In-Reply-To: <CAPP7u0WH9w26Y9ai-EQTTeq7Rz_=7u2-=t4nhHmjwh-UAiyqeQ@mail.gmail.com>
Christian Brauner <christian.brauner@canonical.com> writes:
> On Wed, Apr 25, 2018, 00:41 Eric W. Biederman <ebiederm@xmission.com> wrote:
>
> Bah. This code is obviously correct and probably wrong.
>
> How do we deliver uevents for network devices that are outside of the
> initial user namespace? The kernel still needs to deliver those.
>
> The logic to figure out which network namespace a device needs to be
> delivered to is is present in kobj_bcast_filter. That logic will almost
> certainly need to be turned inside out. Sign not as easy as I would
> have hoped.
>
> My first patch that we discussed put additional filtering logic into kobj_bcast_filter for that very reason. But I can move that logic
> out and come up with a new patch.
I may have mis-understood.
I heard and am still hearing additional filtering to reduce the places
the packet is delievered.
I am saying something needs to change to increase the number of places
the packet is delivered.
For the special class of devices that kobj_bcast_filter would apply to
those need to be delivered to netowrk namespaces that are no longer on
uevent_sock_list.
So the code fundamentally needs to split into two paths. Ordinary
devices that use uevent_sock_list. Network devices that are just
delivered in their own network namespace.
netlink_broadcast_filtered gets to go away completely.
The logic of figuring out the network namespace though becomes trickier.
Now it may make sense to have all of that as an additional patch on top
of this one or perhaps a precursor patch that addresses the problem. We
will unfortunately drop those uevents today because their uids are not
valid. But they are not delivered anywhere else so to allow them to be
received we need to fix them.
Eric
>
> Christian Brauner <christian.brauner@ubuntu.com> writes:
> > commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
> >
> > enabled sending hotplug events into all network namespaces back in 2010.
> > Over time the set of uevents that get sent into all network namespaces has
> > shrunk a little. We have now reached the point where hotplug events for all
> > devices that carry a namespace tag are filtered according to that
> > namespace. Specifically, they are filtered whenever the namespace tag of
> > the kobject does not match the namespace tag of the netlink socket. One
> > example are network devices. Uevents for network devices only show up in
> > the network namespaces these devices are moved to or created in.
> >
> > However, any uevent for a kobject that does not have a namespace tag
> > associated with it will not be filtered and we will broadcast it into all
> > network namespaces. This behavior stopped making sense when user namespaces
> > were introduced.
> >
> > This patch restricts uevents to the initial user namespace for a couple of
> > reasons that have been extensively discusses on the mailing list [1].
> > - Thundering herd:
> > Broadcasting uevents into all network namespaces introduces significant
> > overhead.
> > All processes that listen to uevents running in non-initial user
> > namespaces will end up responding to uevents that will be meaningless to
> > them. Mainly, because non-initial user namespaces cannot easily manage
> > devices unless they have a privileged host-process helping them out. This
> > means that there will be a thundering herd of activity when there
> > shouldn't be any.
> > - Uevents from non-root users are already filtered in userspace:
> > Uevents are filtered by userspace in a user namespace because the
> > received uid != 0. Instead the uid associated with the event will be
> > 65534 == "nobody" because the global root uid is not mapped.
> > This means we can safely and without introducing regressions modify the
> > kernel to not send uevents into all network namespaces whose owning user
> > namespace is not the initial user namespace because we know that
> > userspace will ignore the message because of the uid anyway. I have
> > a) verified that is is true for every udev implementation out there b)
> > that this behavior has been present in all udev implementations from the
> > very beginning.
> > - Removing needless overhead/Increasing performance:
> > Currently, the uevent socket for each network namespace is added to the
> > global variable uevent_sock_list. The list itself needs to be protected
> > by a mutex. So everytime a uevent is generated the mutex is taken on the
> > list. The mutex is held *from the creation of the uevent (memory
> > allocation, string creation etc. until all uevent sockets have been
> > handled*. This is aggravated by the fact that for each uevent socket that
> > has listeners the mc_list must be walked as well which means we're
> > talking O(n^2) here. Given that a standard Linux workload usually has
> > quite a lot of network namespaces and - in the face of containers - a lot
> > of user namespaces this quickly becomes a performance problem (see
> > "Thundering herd" above). By just recording uevent sockets of network
> > namespaces that are owned by the initial user namespace we significantly
> > increase performance in this codepath.
> > - Injecting uevents:
> > There's a valid argument that containers might be interested in receiving
> > device events especially if they are delegated to them by a privileged
> > userspace process. One prime example are SR-IOV enabled devices that are
> > explicitly designed to be handed of to other users such as VMs or
> > containers.
> > This use-case can now be correctly handled since
> > commit 692ec06d7c92 ("netns: send uevent messages"). This commit
> > introduced the ability to send uevents from userspace. As such we can let
> > a sufficiently privileged (CAP_SYS_ADMIN in the owning user namespace of
> > the network namespace of the netlink socket) userspace process make a
> > decision what uevents should be sent. This removes the need to blindly
> > broadcast uevents into all user namespaces and provides a performant and
> > safe solution to this problem.
> > - Filtering logic:
> > This patch filters by *owning user namespace of the network namespace a
> > given task resides in* and not by user namespace of the task per se. This
> > means if the user namespace of a given task is unshared but the network
> > namespace is kept and is owned by the initial user namespace a listener
> > that is opening the uevent socket in that network namespace can still
> > listen to uevents.
> >
> > [1]: https://lkml.org/lkml/2018/4/4/739
> > Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> > ---
> > Changelog v1->v2:
> > * patch unchanged
> > Changelog v0->v1:
> > * patch unchanged
> > ---
> > lib/kobject_uevent.c | 18 ++++++++++++------
> > 1 file changed, 12 insertions(+), 6 deletions(-)
> >
> > diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> > index 15ea216a67ce..f5f5038787ac 100644
> > --- a/lib/kobject_uevent.c
> > +++ b/lib/kobject_uevent.c
> > @@ -703,9 +703,13 @@ static int uevent_net_init(struct net *net)
> >
> > net->uevent_sock = ue_sk;
> >
> > - mutex_lock(&uevent_sock_mutex);
> > - list_add_tail(&ue_sk->list, &uevent_sock_list);
> > - mutex_unlock(&uevent_sock_mutex);
> > + /* Restrict uevents to initial user namespace. */
> > + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> > + mutex_lock(&uevent_sock_mutex);
> > + list_add_tail(&ue_sk->list, &uevent_sock_list);
> > + mutex_unlock(&uevent_sock_mutex);
> > + }
> > +
> > return 0;
> > }
> >
> > @@ -713,9 +717,11 @@ static void uevent_net_exit(struct net *net)
> > {
> > struct uevent_sock *ue_sk = net->uevent_sock;
> >
> > - mutex_lock(&uevent_sock_mutex);
> > - list_del(&ue_sk->list);
> > - mutex_unlock(&uevent_sock_mutex);
> > + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> > + mutex_lock(&uevent_sock_mutex);
> > + list_del(&ue_sk->list);
> > + mutex_unlock(&uevent_sock_mutex);
> > + }
> >
> > netlink_kernel_release(ue_sk->sk);
> > kfree(ue_sk);
^ permalink raw reply
* Re: [PATCH net-next 1/2 v2] netns: restrict uevents
From: Christian Brauner @ 2018-04-24 22:54 UTC (permalink / raw)
To: Eric W. Biederman
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <87po2oz0s8.fsf@xmission.com>
On Tue, Apr 24, 2018 at 05:40:07PM -0500, Eric W. Biederman wrote:
>
> Bah. This code is obviously correct and probably wrong.
>
> How do we deliver uevents for network devices that are outside of the
> initial user namespace? The kernel still needs to deliver those.
>
> The logic to figure out which network namespace a device needs to be
> delivered to is is present in kobj_bcast_filter. That logic will almost
> certainly need to be turned inside out. Sign not as easy as I would
> have hoped.
That's why my initial patch [1] added additional filtering logic to
kobj_bcast_filter(). But since we care about performance improvements as
well I can come up with a patch that moves this logic out of
kobj_bcast_filter().
Christian
[1]: https://www.spinics.net/lists/netdev/msg494487.html
>
> Eric
>
> Christian Brauner <christian.brauner@ubuntu.com> writes:
> > commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
> >
> > enabled sending hotplug events into all network namespaces back in 2010.
> > Over time the set of uevents that get sent into all network namespaces has
> > shrunk a little. We have now reached the point where hotplug events for all
> > devices that carry a namespace tag are filtered according to that
> > namespace. Specifically, they are filtered whenever the namespace tag of
> > the kobject does not match the namespace tag of the netlink socket. One
> > example are network devices. Uevents for network devices only show up in
> > the network namespaces these devices are moved to or created in.
> >
> > However, any uevent for a kobject that does not have a namespace tag
> > associated with it will not be filtered and we will broadcast it into all
> > network namespaces. This behavior stopped making sense when user namespaces
> > were introduced.
> >
> > This patch restricts uevents to the initial user namespace for a couple of
> > reasons that have been extensively discusses on the mailing list [1].
> > - Thundering herd:
> > Broadcasting uevents into all network namespaces introduces significant
> > overhead.
> > All processes that listen to uevents running in non-initial user
> > namespaces will end up responding to uevents that will be meaningless to
> > them. Mainly, because non-initial user namespaces cannot easily manage
> > devices unless they have a privileged host-process helping them out. This
> > means that there will be a thundering herd of activity when there
> > shouldn't be any.
> > - Uevents from non-root users are already filtered in userspace:
> > Uevents are filtered by userspace in a user namespace because the
> > received uid != 0. Instead the uid associated with the event will be
> > 65534 == "nobody" because the global root uid is not mapped.
> > This means we can safely and without introducing regressions modify the
> > kernel to not send uevents into all network namespaces whose owning user
> > namespace is not the initial user namespace because we know that
> > userspace will ignore the message because of the uid anyway. I have
> > a) verified that is is true for every udev implementation out there b)
> > that this behavior has been present in all udev implementations from the
> > very beginning.
> > - Removing needless overhead/Increasing performance:
> > Currently, the uevent socket for each network namespace is added to the
> > global variable uevent_sock_list. The list itself needs to be protected
> > by a mutex. So everytime a uevent is generated the mutex is taken on the
> > list. The mutex is held *from the creation of the uevent (memory
> > allocation, string creation etc. until all uevent sockets have been
> > handled*. This is aggravated by the fact that for each uevent socket that
> > has listeners the mc_list must be walked as well which means we're
> > talking O(n^2) here. Given that a standard Linux workload usually has
> > quite a lot of network namespaces and - in the face of containers - a lot
> > of user namespaces this quickly becomes a performance problem (see
> > "Thundering herd" above). By just recording uevent sockets of network
> > namespaces that are owned by the initial user namespace we significantly
> > increase performance in this codepath.
> > - Injecting uevents:
> > There's a valid argument that containers might be interested in receiving
> > device events especially if they are delegated to them by a privileged
> > userspace process. One prime example are SR-IOV enabled devices that are
> > explicitly designed to be handed of to other users such as VMs or
> > containers.
> > This use-case can now be correctly handled since
> > commit 692ec06d7c92 ("netns: send uevent messages"). This commit
> > introduced the ability to send uevents from userspace. As such we can let
> > a sufficiently privileged (CAP_SYS_ADMIN in the owning user namespace of
> > the network namespace of the netlink socket) userspace process make a
> > decision what uevents should be sent. This removes the need to blindly
> > broadcast uevents into all user namespaces and provides a performant and
> > safe solution to this problem.
> > - Filtering logic:
> > This patch filters by *owning user namespace of the network namespace a
> > given task resides in* and not by user namespace of the task per se. This
> > means if the user namespace of a given task is unshared but the network
> > namespace is kept and is owned by the initial user namespace a listener
> > that is opening the uevent socket in that network namespace can still
> > listen to uevents.
> >
> > [1]: https://lkml.org/lkml/2018/4/4/739
> > Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> > ---
> > Changelog v1->v2:
> > * patch unchanged
> > Changelog v0->v1:
> > * patch unchanged
> > ---
> > lib/kobject_uevent.c | 18 ++++++++++++------
> > 1 file changed, 12 insertions(+), 6 deletions(-)
> >
> > diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> > index 15ea216a67ce..f5f5038787ac 100644
> > --- a/lib/kobject_uevent.c
> > +++ b/lib/kobject_uevent.c
> > @@ -703,9 +703,13 @@ static int uevent_net_init(struct net *net)
> >
> > net->uevent_sock = ue_sk;
> >
> > - mutex_lock(&uevent_sock_mutex);
> > - list_add_tail(&ue_sk->list, &uevent_sock_list);
> > - mutex_unlock(&uevent_sock_mutex);
> > + /* Restrict uevents to initial user namespace. */
> > + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> > + mutex_lock(&uevent_sock_mutex);
> > + list_add_tail(&ue_sk->list, &uevent_sock_list);
> > + mutex_unlock(&uevent_sock_mutex);
> > + }
> > +
> > return 0;
> > }
> >
> > @@ -713,9 +717,11 @@ static void uevent_net_exit(struct net *net)
> > {
> > struct uevent_sock *ue_sk = net->uevent_sock;
> >
> > - mutex_lock(&uevent_sock_mutex);
> > - list_del(&ue_sk->list);
> > - mutex_unlock(&uevent_sock_mutex);
> > + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> > + mutex_lock(&uevent_sock_mutex);
> > + list_del(&ue_sk->list);
> > + mutex_unlock(&uevent_sock_mutex);
> > + }
> >
> > netlink_kernel_release(ue_sk->sk);
> > kfree(ue_sk);
^ permalink raw reply
* Re: [PATCH net-next 2/2 v2] netns: isolate seqnums to use per-netns locks
From: Eric W. Biederman @ 2018-04-24 22:52 UTC (permalink / raw)
To: Christian Brauner
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <20180424222049.GA21073@gmail.com>
Christian Brauner <christian.brauner@canonical.com> writes:
> On Tue, Apr 24, 2018 at 04:52:20PM -0500, Eric W. Biederman wrote:
>> Christian Brauner <christian.brauner@ubuntu.com> writes:
>>
>> > Now that it's possible to have a different set of uevents in different
>> > network namespaces, per-network namespace uevent sequence numbers are
>> > introduced. This increases performance as locking is now restricted to the
>> > network namespace affected by the uevent rather than locking everything.
>> > Testing revealed significant performance improvements. For details see
>> > "Testing" below.
>>
>> Maybe. Your locking is wrong, and a few other things are wrong. see
>> below.
>
> Thanks for the review! Happy to rework this until it's in a mergeable shape.
>
>>
>> > Since commit 692ec06 ("netns: send uevent messages") network namespaces not
>> > owned by the intial user namespace can be sent uevents from a sufficiently
>> > privileged userspace process.
>> > In order to send a uevent into a network namespace not owned by the initial
>> > user namespace we currently still need to take the *global mutex* that
>> > locks the uevent socket list even though the list *only contains network
>> > namespaces owned by the initial user namespace*. This needs to be done
>> > because the uevent counter is a global variable. Taking the global lock is
>> > performance sensitive since a user on the host can spawn a pool of n
>> > process that each create their own new user and network namespaces and then
>> > go on to inject uevents in parallel into the network namespace of all of
>> > these processes. This can have a significant performance impact for the
>> > host's udevd since it means that there can be a lot of delay between a
>> > device being added and the corresponding uevent being sent out and
>> > available for processing by udevd. It also means that each network
>> > namespace not owned by the initial user namespace which userspace has sent
>> > a uevent to will need to wait until the lock becomes available.
>> >
>> > Implementation:
>> > This patch gives each network namespace its own uevent sequence number.
>> > Each network namespace not owned by the initial user namespace receives its
>> > own mutex. The struct uevent_sock is opaque to callers outside of kobject.c
>> > so the mutex *can* and *is* only ever accessed in lib/kobject.c. In this
>> > file it is clearly documented which lock has to be taken. All network
>> > namespaces owned by the initial user namespace will still share the same
>> > lock since they are all served sequentially via the uevent socket list.
>> > This decouples the locking and ensures that the host retrieves uevents as
>> > fast as possible even if there are a lot of uevents injected into network
>> > namespaces not owned by the initial user namespace. In addition, each
>> > network namespace not owned by the initial user namespace does not have to
>> > wait on any other network namespace not sharing the same user namespace.
>> >
>> > Testing:
>> > Two 4.17-rc1 test kernels were compiled. One with per netns uevent seqnums
>> > with decoupled locking and one without. To ensure that testing made sense
>> > both kernels carried the patch to remove network namespaces not owned by
>> > the initial user namespace from the uevent socket list.
>> > Three tests were constructed. All of them showed significant performance
>> > improvements with per-netns uevent sequence numbers and decoupled locking.
>> >
>> > # Testcase 1:
>> > Only Injecting Uevents into network namespaces not owned by the initial
>> > user namespace.
>> > - created 1000 new user namespace + network namespace pairs
>> > - opened a uevent listener in each of those namespace pairs
>> > - injected uevents into each of those network namespaces 10,000 times
>> > meaning 10,000,000 (10 million) uevents were injected. (The high
>> > number of uevent injections should get rid of a lot of jitter.)
>> > The injection was done by fork()ing 1000 uevent injectors in a simple
>> > for-loop to ensure that uevents were injected in parallel.
>> > - mean transaction time was calculated:
>> > - *without* uevent sequence number namespacing: 67 μs
>> > - *with* uevent sequence number namespacing: 55 μs
>> > - makes a difference of: 12 μs
>> > - a t-test was performed on the two data vectors which revealed
>> > shows significant performance improvements:
>> > Welch Two Sample t-test
>> > data: x1 and y1
>> > t = 405.16, df = 18883000, p-value < 2.2e-16
>> > alternative hypothesis: true difference in means is not equal to 0
>> > 95 percent confidence interval:
>> > 12.14949 12.26761
>> > sample estimates:
>> > mean of x mean of y
>> > 68.48594 56.27739
>> >
>> > # Testcase 2:
>> > Injecting Uevents into network namespaces not owned by the initial user
>> > namespace and network namespaces owned by the initial user namespace.
>> > - created 500 new user namespace + network namespace pairs
>> > - created 500 new network namespace pairs
>> > - opened a uevent listener in each of those namespace pairs
>> > - injected uevents into each of those network namespaces 10,000 times
>> > meaning 10,000,000 (10 million) uevents were injected. (The high
>> > number of uevent injections should get rid of a lot of jitter.)
>> > The injection was done by fork()ing 1000 uevent injectors in a simple
>> > for-loop to ensure that uevents were injected in parallel.
>> > - mean transaction time was calculated:
>> > - *without* uevent sequence number namespacing: 572 μs
>> > - *with* uevent sequence number namespacing: 514 μs
>> > - makes a difference of: 58 μs
>> > - a t-test was performed on the two data vectors which revealed
>> > shows significant performance improvements:
>> > Welch Two Sample t-test
>> > data: x2 and y2
>> > t = 38.685, df = 19682000, p-value < 2.2e-16
>> > alternative hypothesis: true difference in means is not equal to 0
>> > 95 percent confidence interval:
>> > 55.10630 60.98815
>> > sample estimates:
>> > mean of x mean of y
>> > 572.9684 514.9211
>> >
>> > # Testcase 3:
>> > Created 500 new user namespace + network namespace pairs *without uevent
>> > listeners*
>> > - created 500 new network namespace pairs *without uevent listeners*
>> > - injected uevents into each of those network namespaces 10,000 times
>> > meaning 10,000,000 (10 million) uevents were injected. (The high number
>> > of uevent injections should get rid of a lot of jitter.)
>> > The injection was done by fork()ing 1000 uevent injectors in a simple
>> > for-loop to ensure that uevents were injected in parallel.
>> > - mean transaction time was calculated:
>> > - *without* uevent sequence number namespacing: 206 μs
>> > - *with* uevent sequence number namespacing: 163 μs
>> > - makes a difference of: 43 μs
>> > - a t-test was performed on the two data vectors which revealed
>> > shows significant performance improvements:
>> > Welch Two Sample t-test
>> > data: x3 and y3
>> > t = 58.37, df = 17711000, p-value < 2.2e-16
>> > alternative hypothesis: true difference in means is not equal to 0
>> > 95 percent confidence interval:
>> > 41.77860 44.68178
>> > sample estimates:
>> > mean of x mean of y
>> > 207.2632 164.0330
>> >
>> > Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
>> > ---
>> > Changelog v1->v2:
>> > * non-functional change: fix indendation for C directives in
>> > kernel/ksysfs.c
>> > Changelog v0->v1:
>> > * add detailed test results to the commit message
>> > * account for kernels compiled without CONFIG_NET
>> > ---
>> > include/linux/kobject.h | 2 +
>> > include/net/net_namespace.h | 3 ++
>> > kernel/ksysfs.c | 11 +++-
>> > lib/kobject_uevent.c | 104 +++++++++++++++++++++++++++++-------
>> > net/core/net_namespace.c | 14 +++++
>> > 5 files changed, 114 insertions(+), 20 deletions(-)
>> >
>> > diff --git a/include/linux/kobject.h b/include/linux/kobject.h
>> > index 7f6f93c3df9c..4e608968907f 100644
>> > --- a/include/linux/kobject.h
>> > +++ b/include/linux/kobject.h
>> > @@ -36,8 +36,10 @@
>> > extern char uevent_helper[];
>> > #endif
>> >
>> > +#ifndef CONFIG_NET
>> > /* counter to tag the uevent, read only except for the kobject core */
>> > extern u64 uevent_seqnum;
>> > +#endif
>>
>> That smells like an implementation bug somewhere.
>
> Sorry, I'm not following. I'm I'm not mistaken there won't be any struct
> net when CONFIG_NET=n. This has been reported by kbuild robot with alpha
> and CONFIG_NET=n.
The Kbuild robot isn't wrong. I am just saying this likely comes up
because the code is not clean.
If uevent_seqnum is tied to a network namespace than it doesn't make
sense to exist without netork namespaces.
At the very least this feels like an untested special case. Untested
special cases tend to break when people are not paying attention which
can be dangerous.
>> > /*
>> > * The actions here must match the index to the string array
>> > diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
>> > index 47e35cce3b64..e4e171b1ba69 100644
>> > --- a/include/net/net_namespace.h
>> > +++ b/include/net/net_namespace.h
>> > @@ -85,6 +85,8 @@ struct net {
>> > struct sock *genl_sock;
>> >
>> > struct uevent_sock *uevent_sock; /* uevent socket */
>> > + /* counter to tag the uevent, read only except for the kobject core */
>> > + u64 uevent_seqnum;
>> >
>> > struct list_head dev_base_head;
>> > struct hlist_head *dev_name_head;
>> > @@ -189,6 +191,7 @@ extern struct list_head net_namespace_list;
>> >
>> > struct net *get_net_ns_by_pid(pid_t pid);
>> > struct net *get_net_ns_by_fd(int fd);
>> > +u64 get_ns_uevent_seqnum_by_vpid(void);
>> >
>> > #ifdef CONFIG_SYSCTL
>> > void ipx_register_sysctl(void);
>> > diff --git a/kernel/ksysfs.c b/kernel/ksysfs.c
>> > index 46ba853656f6..38b70b90a21f 100644
>> > --- a/kernel/ksysfs.c
>> > +++ b/kernel/ksysfs.c
>> > @@ -19,6 +19,7 @@
>> > #include <linux/sched.h>
>> > #include <linux/capability.h>
>> > #include <linux/compiler.h>
>> > +#include <net/net_namespace.h>
>> >
>> > #include <linux/rcupdate.h> /* rcu_expedited and rcu_normal */
>> >
>> > @@ -33,7 +34,15 @@ static struct kobj_attribute _name##_attr = \
>> > static ssize_t uevent_seqnum_show(struct kobject *kobj,
>> > struct kobj_attribute *attr, char *buf)
>> > {
>> > - return sprintf(buf, "%llu\n", (unsigned long long)uevent_seqnum);
>> > + u64 seqnum;
>> > +
>> > +#ifdef CONFIG_NET
>> > + seqnum = get_ns_uevent_seqnum_by_vpid();
>> > +#else
>> > + seqnum = uevent_seqnum;
>> > +#endif
>>
>> This can be simplified to be just:
>> seqnum = current->nsproxy->net_ns->uevent_seqnum;
>
> Does that work even if CONFIG_NET=n?
No. It is a NULL pointer dereference but the net_ns member continues
to exist.
>> Except that is not correct either. As every instance of sysfs has a
>> network namespace associated with it, and you are not fetching that
>> network namespace.
>
> I'm not yet familiar with all aspects of sysfs so thanks for pointing
> that out. Then I'll try to come up with a way to fetch the network
> namespace associated with sysfs. Unless you already know exactly how to
> do this and can point it out.
> This would also lets us drop get_ns_uevent_seqnum_by_vpid().
Everywhere else the distinction has been at a directory level. With one
the set of directory entries being different depending on which network
namespace sysfs was mounted in. For /sys/kernel/seqnum that looks like
fresh work.
Plus I expect there are some embedded systems without networking that
may still need uevents via /sbin/hotplug. Which makes this doubly
tricky.
I honestly don't know the answer to this one.
>> Typically this would call for making this file per network namespace
>> so you would have this information available. Sigh. I don't know if
>> there is an easy way to do that for this file.
>>
>> > +
>> > + return sprintf(buf, "%llu\n", (unsigned long long)seqnum);
>> > }
>> > KERNEL_ATTR_RO(uevent_seqnum);
>> >
>> > diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
>> > index f5f5038787ac..5da20def556d 100644
>> > --- a/lib/kobject_uevent.c
>> > +++ b/lib/kobject_uevent.c
>> > @@ -29,21 +29,42 @@
>> > #include <net/net_namespace.h>
>> >
>> >
>> > +#ifndef CONFIG_NET
>> > u64 uevent_seqnum;
>> > +#endif
>> > +
>> > #ifdef CONFIG_UEVENT_HELPER
>> > char uevent_helper[UEVENT_HELPER_PATH_LEN] = CONFIG_UEVENT_HELPER_PATH;
>> > #endif
>> >
>> > +/*
>> > + * Size a buffer needs to be in order to hold the largest possible sequence
>> > + * number stored in a u64 including \0 byte: 2^64 - 1 = 21 chars.
>> > + */
>> > +#define SEQNUM_BUFSIZE (sizeof("SEQNUM=") + 21)
>> > struct uevent_sock {
>> > struct list_head list;
>> > struct sock *sk;
>> > + /*
>> > + * This mutex protects uevent sockets and the uevent counter of
>> > + * network namespaces *not* owned by init_user_ns.
>> > + * For network namespaces owned by init_user_ns this lock is *not*
>> > + * valid instead the global uevent_sock_mutex must be used!
>> > + */
>> > + struct mutex sk_mutex;
>> > };
>> >
>> > #ifdef CONFIG_NET
>> > static LIST_HEAD(uevent_sock_list);
>> > #endif
>> >
>> > -/* This lock protects uevent_seqnum and uevent_sock_list */
>> > +/*
>> > + * This mutex protects uevent sockets and the uevent counter of network
>> > + * namespaces owned by init_user_ns.
>> > + * For network namespaces not owned by init_user_ns this lock is *not*
>> > + * valid instead the network namespace specific sk_mutex in struct
>> > + * uevent_sock must be used!
>> > + */
>> > static DEFINE_MUTEX(uevent_sock_mutex);
>> >
>> > /* the strings here must match the enum in include/linux/kobject.h */
>> > @@ -253,6 +274,22 @@ static int kobj_bcast_filter(struct sock *dsk, struct sk_buff *skb, void *data)
>> >
>> > return 0;
>> > }
>> > +
>> > +static bool can_hold_seqnum(const struct kobj_uevent_env *env, size_t len)
>> > +{
>> > + if (env->envp_idx >= ARRAY_SIZE(env->envp)) {
>> > + WARN(1, KERN_ERR "Failed to append sequence number. "
>> > + "Too many uevent variables\n");
>> > + return false;
>> > + }
>> > +
>> > + if ((env->buflen + len) > UEVENT_BUFFER_SIZE) {
>> > + WARN(1, KERN_ERR "Insufficient space to append sequence number\n");
>> > + return false;
>> > + }
>> > +
>> > + return true;
>> > +}
>> > #endif
>> >
>> > #ifdef CONFIG_UEVENT_HELPER
>> > @@ -308,18 +345,22 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
>> >
>> > /* send netlink message */
>> > list_for_each_entry(ue_sk, &uevent_sock_list, list) {
>> > + /* bump sequence number */
>> > + u64 seqnum = ++sock_net(ue_sk->sk)->uevent_seqnum;
>> > struct sock *uevent_sock = ue_sk->sk;
>> > + char buf[SEQNUM_BUFSIZE];
>> >
>> > if (!netlink_has_listeners(uevent_sock, 1))
>> > continue;
>> >
>> > if (!skb) {
>> > - /* allocate message with the maximum possible size */
>> > + /* calculate header length */
>> > size_t len = strlen(action_string) + strlen(devpath) + 2;
>> > char *scratch;
>> >
>> > + /* allocate message with the maximum possible size */
>> > retval = -ENOMEM;
>> > - skb = alloc_skb(len + env->buflen, GFP_KERNEL);
>> > + skb = alloc_skb(len + env->buflen + SEQNUM_BUFSIZE, GFP_KERNEL);
>> > if (!skb)
>> > continue;
>> >
>> > @@ -327,11 +368,24 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
>> > scratch = skb_put(skb, len);
>> > sprintf(scratch, "%s@%s", action_string, devpath);
>> >
>> > + /* add env */
>> > skb_put_data(skb, env->buf, env->buflen);
>> >
>> > NETLINK_CB(skb).dst_group = 1;
>> > }
>> >
>> > + /* prepare netns seqnum */
>> > + retval = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu", seqnum);
>> > + if (retval < 0 || retval >= SEQNUM_BUFSIZE)
>> > + continue;
>> > + retval++;
>> > +
>> > + if (!can_hold_seqnum(env, retval))
>> > + continue;
>>
>> You have allocated enough space in the skb why does can_hold_seqnum make
>> sense?
>
> Because it doesn't check whether the socket buffer can hold the sequence
> number but checks whether the uevent buffer size in "env" can hold it.
> uevents are only delivered if the env buffer is large enough to hold all
> of the info including the sequence number. That's independent of the
> socket buffer.
The practical question is why does that get called every time through
the loop?
>>
>> Do you need to back seqnum out of the env later for this to work twice
>> in a row?
>
> I guess I can just override it. It just felt cleaner to trim it.
You trim it from the socket, but what about the environment?
>
>>
>> > +
>> > + /* append netns seqnum */
>> > + skb_put_data(skb, buf, retval);
>> > +
>> > retval = netlink_broadcast_filtered(uevent_sock, skb_get(skb),
>> > 0, 1, GFP_KERNEL,
>> > kobj_bcast_filter,
>> > @@ -339,8 +393,13 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
>> > /* ENOBUFS should be handled in userspace */
>> > if (retval == -ENOBUFS || retval == -ESRCH)
>> > retval = 0;
>> > +
>> > + /* remove netns seqnum */
>> > + skb_trim(skb, env->buflen);
>>
>> Have you checked to see if the seqnum actually makes it to userspace.
>
> Yes, I did. I also wonder why it wouldn't make it. Any specific reason
> why you suspect this?
trim? I am not really certain what that does and if when you deliever
to multiple network namespaces if it might not cause the wrong seqnum
to be delivered. You are using the same initial buffer.
I haven't stared at the socket buffer code enough to remember off hand
what the semantics are. Perhaps whatever netlink_broadcast does
is sufficient to prevent issues.
>
>> > }
>> > consume_skb(skb);
>> > +#else
>> > + uevent_seqnum++;
>> > #endif
>> > return retval;
>> > }
>> > @@ -510,14 +569,7 @@ int kobject_uevent_env(struct kobject *kobj, enum kobject_action action,
>> > }
>> >
>> > mutex_lock(&uevent_sock_mutex);
>> > - /* we will send an event, so request a new sequence number */
>> > - retval = add_uevent_var(env, "SEQNUM=%llu", (unsigned long long)++uevent_seqnum);
>> > - if (retval) {
>> > - mutex_unlock(&uevent_sock_mutex);
>> > - goto exit;
>> > - }
>> > - retval = kobject_uevent_net_broadcast(kobj, env, action_string,
>> > - devpath);
>> > + retval = kobject_uevent_net_broadcast(kobj, env, action_string, devpath);
>> > mutex_unlock(&uevent_sock_mutex);
>>
>> How does all of this work with events for network devices that are not
>> in the initial network namespace. This looks to me like this code fails
>> to take the sk_mutex.
>
> But in this list only non-initial network namespaces that are owned by
> the initial user namespace are recorded and for these uevent_sock_mutex
> has to be taken. Am I missing something?
The bug I missed in your first patch. We still have to deliver uevents
to secondary network namespaces for their network devices.
At which point you need your new sk_mutex.
>> > #ifdef CONFIG_UEVENT_HELPER
>> > @@ -605,17 +657,18 @@ int add_uevent_var(struct kobj_uevent_env *env, const char *format, ...)
>> > EXPORT_SYMBOL_GPL(add_uevent_var);
>> >
>> > #if defined(CONFIG_NET)
>> > -static int uevent_net_broadcast(struct sock *usk, struct sk_buff *skb,
>> > +static int uevent_net_broadcast(struct uevent_sock *ue_sk, struct sk_buff *skb,
>> > struct netlink_ext_ack *extack)
>> > {
>> > - /* u64 to chars: 2^64 - 1 = 21 chars */
>> > - char buf[sizeof("SEQNUM=") + 21];
>> > + struct sock *usk = ue_sk->sk;
>> > + char buf[SEQNUM_BUFSIZE];
>> > struct sk_buff *skbc;
>> > int ret;
>> >
>> > /* bump and prepare sequence number */
>> > - ret = snprintf(buf, sizeof(buf), "SEQNUM=%llu", ++uevent_seqnum);
>> > - if (ret < 0 || (size_t)ret >= sizeof(buf))
>> > + ret = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu",
>> > + ++sock_net(ue_sk->sk)->uevent_seqnum);
>> > + if (ret < 0 || ret >= SEQNUM_BUFSIZE)
>> > return -ENOMEM;
>> > ret++;
>> >
>> > @@ -668,9 +721,15 @@ static int uevent_net_rcv_skb(struct sk_buff *skb, struct nlmsghdr *nlh,
>> > return -EPERM;
>> > }
>> >
>> > - mutex_lock(&uevent_sock_mutex);
>> > - ret = uevent_net_broadcast(net->uevent_sock->sk, skb, extack);
>> > - mutex_unlock(&uevent_sock_mutex);
>> > + if (net->user_ns == &init_user_ns)
>> > + mutex_lock(&uevent_sock_mutex);
>> > + else
>> > + mutex_lock(&net->uevent_sock->sk_mutex);
>> > + ret = uevent_net_broadcast(net->uevent_sock, skb, extack);
>> > + if (net->user_ns == &init_user_ns)
>> > + mutex_unlock(&uevent_sock_mutex);
>> > + else
>> > + mutex_unlock(&net->uevent_sock->sk_mutex);
>> >
>> > return ret;
>> > }
>> > @@ -708,6 +767,13 @@ static int uevent_net_init(struct net *net)
>> > mutex_lock(&uevent_sock_mutex);
>> > list_add_tail(&ue_sk->list, &uevent_sock_list);
>> > mutex_unlock(&uevent_sock_mutex);
>> > + } else {
>> > + /*
>> > + * Uevent sockets and counters for network namespaces
>> > + * not owned by the initial user namespace have their
>> > + * own mutex.
>> > + */
>> > + mutex_init(&ue_sk->sk_mutex);
>> > }
>> >
>> > return 0;
>> > diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
>> > index a11e03f920d3..8894638f5150 100644
>> > --- a/net/core/net_namespace.c
>> > +++ b/net/core/net_namespace.c
>> > @@ -618,6 +618,20 @@ struct net *get_net_ns_by_pid(pid_t pid)
>> > }
>> > EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
>> >
>> > +u64 get_ns_uevent_seqnum_by_vpid(void)
>> > +{
>> > + pid_t cur_pid;
>> > + struct net *net;
>> > +
>> > + cur_pid = task_pid_vnr(current);
>> > + net = get_net_ns_by_pid(cur_pid);
>> > + if (IS_ERR(net))
>> > + return 0;
>> > +
>> > + return net->uevent_seqnum;
>> > +}
>> > +EXPORT_SYMBOL_GPL(get_ns_uevent_seqnum_by_vpid);
>>
>> I just have to say this function is completely crazy.
>> You go from the tsk to the pid back to the tsk.
>> And you leak a struct net pointer.
>>
>> It is much simpler and less racy to say:
>>
>> current->nsproxy->net_ns->uevent_seqnum;
>>
>> That you are accessing current->nsproxy means nsproxy can't
>> change. The rcu_read_lock etc that get_net_ns_by_pid does
>> is there for accessing non-current tasks.
>>
>>
>>
>> > static __net_init int net_ns_net_init(struct net *net)
>> > {
>> > #ifdef CONFIG_NET_NS
^ permalink raw reply
* Re: [PATCH net-next 1/2 v2] netns: restrict uevents
From: Eric W. Biederman @ 2018-04-24 22:40 UTC (permalink / raw)
To: Christian Brauner
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <20180424204335.12904-2-christian.brauner@ubuntu.com>
Bah. This code is obviously correct and probably wrong.
How do we deliver uevents for network devices that are outside of the
initial user namespace? The kernel still needs to deliver those.
The logic to figure out which network namespace a device needs to be
delivered to is is present in kobj_bcast_filter. That logic will almost
certainly need to be turned inside out. Sign not as easy as I would
have hoped.
Eric
Christian Brauner <christian.brauner@ubuntu.com> writes:
> commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
>
> enabled sending hotplug events into all network namespaces back in 2010.
> Over time the set of uevents that get sent into all network namespaces has
> shrunk a little. We have now reached the point where hotplug events for all
> devices that carry a namespace tag are filtered according to that
> namespace. Specifically, they are filtered whenever the namespace tag of
> the kobject does not match the namespace tag of the netlink socket. One
> example are network devices. Uevents for network devices only show up in
> the network namespaces these devices are moved to or created in.
>
> However, any uevent for a kobject that does not have a namespace tag
> associated with it will not be filtered and we will broadcast it into all
> network namespaces. This behavior stopped making sense when user namespaces
> were introduced.
>
> This patch restricts uevents to the initial user namespace for a couple of
> reasons that have been extensively discusses on the mailing list [1].
> - Thundering herd:
> Broadcasting uevents into all network namespaces introduces significant
> overhead.
> All processes that listen to uevents running in non-initial user
> namespaces will end up responding to uevents that will be meaningless to
> them. Mainly, because non-initial user namespaces cannot easily manage
> devices unless they have a privileged host-process helping them out. This
> means that there will be a thundering herd of activity when there
> shouldn't be any.
> - Uevents from non-root users are already filtered in userspace:
> Uevents are filtered by userspace in a user namespace because the
> received uid != 0. Instead the uid associated with the event will be
> 65534 == "nobody" because the global root uid is not mapped.
> This means we can safely and without introducing regressions modify the
> kernel to not send uevents into all network namespaces whose owning user
> namespace is not the initial user namespace because we know that
> userspace will ignore the message because of the uid anyway. I have
> a) verified that is is true for every udev implementation out there b)
> that this behavior has been present in all udev implementations from the
> very beginning.
> - Removing needless overhead/Increasing performance:
> Currently, the uevent socket for each network namespace is added to the
> global variable uevent_sock_list. The list itself needs to be protected
> by a mutex. So everytime a uevent is generated the mutex is taken on the
> list. The mutex is held *from the creation of the uevent (memory
> allocation, string creation etc. until all uevent sockets have been
> handled*. This is aggravated by the fact that for each uevent socket that
> has listeners the mc_list must be walked as well which means we're
> talking O(n^2) here. Given that a standard Linux workload usually has
> quite a lot of network namespaces and - in the face of containers - a lot
> of user namespaces this quickly becomes a performance problem (see
> "Thundering herd" above). By just recording uevent sockets of network
> namespaces that are owned by the initial user namespace we significantly
> increase performance in this codepath.
> - Injecting uevents:
> There's a valid argument that containers might be interested in receiving
> device events especially if they are delegated to them by a privileged
> userspace process. One prime example are SR-IOV enabled devices that are
> explicitly designed to be handed of to other users such as VMs or
> containers.
> This use-case can now be correctly handled since
> commit 692ec06d7c92 ("netns: send uevent messages"). This commit
> introduced the ability to send uevents from userspace. As such we can let
> a sufficiently privileged (CAP_SYS_ADMIN in the owning user namespace of
> the network namespace of the netlink socket) userspace process make a
> decision what uevents should be sent. This removes the need to blindly
> broadcast uevents into all user namespaces and provides a performant and
> safe solution to this problem.
> - Filtering logic:
> This patch filters by *owning user namespace of the network namespace a
> given task resides in* and not by user namespace of the task per se. This
> means if the user namespace of a given task is unshared but the network
> namespace is kept and is owned by the initial user namespace a listener
> that is opening the uevent socket in that network namespace can still
> listen to uevents.
>
> [1]: https://lkml.org/lkml/2018/4/4/739
> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> ---
> Changelog v1->v2:
> * patch unchanged
> Changelog v0->v1:
> * patch unchanged
> ---
> lib/kobject_uevent.c | 18 ++++++++++++------
> 1 file changed, 12 insertions(+), 6 deletions(-)
>
> diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> index 15ea216a67ce..f5f5038787ac 100644
> --- a/lib/kobject_uevent.c
> +++ b/lib/kobject_uevent.c
> @@ -703,9 +703,13 @@ static int uevent_net_init(struct net *net)
>
> net->uevent_sock = ue_sk;
>
> - mutex_lock(&uevent_sock_mutex);
> - list_add_tail(&ue_sk->list, &uevent_sock_list);
> - mutex_unlock(&uevent_sock_mutex);
> + /* Restrict uevents to initial user namespace. */
> + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> + mutex_lock(&uevent_sock_mutex);
> + list_add_tail(&ue_sk->list, &uevent_sock_list);
> + mutex_unlock(&uevent_sock_mutex);
> + }
> +
> return 0;
> }
>
> @@ -713,9 +717,11 @@ static void uevent_net_exit(struct net *net)
> {
> struct uevent_sock *ue_sk = net->uevent_sock;
>
> - mutex_lock(&uevent_sock_mutex);
> - list_del(&ue_sk->list);
> - mutex_unlock(&uevent_sock_mutex);
> + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> + mutex_lock(&uevent_sock_mutex);
> + list_del(&ue_sk->list);
> + mutex_unlock(&uevent_sock_mutex);
> + }
>
> netlink_kernel_release(ue_sk->sk);
> kfree(ue_sk);
^ permalink raw reply
* Re: [bpf-next PATCH 0/4] selftests for BPF sockmap use cases
From: Daniel Borkmann @ 2018-04-24 22:22 UTC (permalink / raw)
To: John Fastabend, ast; +Cc: netdev
In-Reply-To: <20180423212506.20332.6762.stgit@john-Precision-Tower-5810>
On 04/23/2018 11:30 PM, John Fastabend wrote:
> This series moves ./samples/sockmap into BPF selftests. There are a
> few good reasons to do this. First, by pushing this into selftests
> the tests will be run automatically. Second, sockmap was not really
> a sample of anything anymore, but rather a large set of tests.
>
> Note: There are three recent fixes outstanding against bpf branch
> that can be detected occosionally by the automated tests here.
>
> https://patchwork.ozlabs.org/patch/903138/
> https://patchwork.ozlabs.org/patch/903139/
> https://patchwork.ozlabs.org/patch/903140/
>
> ---
>
> John Fastabend (4):
> bpf: sockmap, code sockmap_test in C
> bpf: sockmap, add a set of tests to run by default
> bpf: sockmap, add selftests
> bpf: sockmap, remove samples program
Applied to bpf-next, thanks John!
^ permalink raw reply
* Re: [pci PATCH v8 0/4] Add support for unmanaged SR-IOV
From: Alexander Duyck @ 2018-04-24 22:22 UTC (permalink / raw)
To: Bjorn Helgaas
Cc: Alexander Duyck, Bjorn Helgaas, linux-pci, virtio-dev, kvm,
Netdev, Daly, Dan, LKML, linux-nvme, Keith Busch, netanel,
Don Dutile, Maximilian Heyne, Wang, Liang-min, Rustad, Mark D,
David Woodhouse, Christoph Hellwig, dwmw
In-Reply-To: <20180424215150.GB72698@bhelgaas-glaptop.roam.corp.google.com>
On Tue, Apr 24, 2018 at 2:51 PM, Bjorn Helgaas <helgaas@kernel.org> wrote:
> On Sat, Apr 21, 2018 at 05:22:27PM -0700, Alexander Duyck wrote:
>> On Sat, Apr 21, 2018 at 1:34 PM, Bjorn Helgaas <helgaas@kernel.org> wrote:
>
>> > For example, I'm not sure what you mean by "devices where the PF is
>> > not capable of managing VF resources."
>> >
>> > It *sounds* like you're saying the hardware works differently on some
>> > devices, but I don't think that's what you mean. I think you're
>> > saying something about which drivers are used for the PF and the VF.
>>
>> That is sort of what I am saying.
>>
>> So for example with ixgbe there is functionality which is controlled
>> in the MMIO space of the PF that affects the functionality of the VFs
>> that are generated on the device. The PF has to rearrange the
>> resources such as queues and interrupts on the device before it can
>> enable SR-IOV, and it could alter those later to limit what the VF is
>> capable of doing.
>>
>> The model I am dealing with via this patch set has a PF that is not
>> much different than the VFs other than the fact that it has some
>> extended configuration space bits in place for SR-IOV, ARI, ACS, and
>> whatever other bits are needed in order to support spawning isolated
>> VFs.
>
> OK, thanks for the explanation, I think I understand what's going on
> now, correct me if I'm mistaken. I added a hint about "PF" for Randy,
> too.
>
> These are on pci/virtualization for v4.18.
I reviewed them and all of the changes to patches 1 & 2 both below,
and in the tree look good to me.
Thanks for taking care of all this.
- Alex
> commit 8effc395c209
> Author: Alexander Duyck <alexander.h.duyck@intel.com>
> Date: Sat Apr 21 15:23:09 2018 -0500
>
> PCI/IOV: Add pci_sriov_configure_simple()
>
> SR-IOV (Single Root I/O Virtualization) is an optional PCIe capability (see
> PCIe r4.0, sec 9). A PCIe Function with the SR-IOV capability is referred
> to as a PF (Physical Function). If SR-IOV is enabled on the PF, several
> VFs (Virtual Functions) may be created. The VFs can be individually
> assigned to virtual machines, which allows them to share a single hardware
> device while being isolated from each other.
>
> Some SR-IOV devices have resources such as queues and interrupts that must
> be set up in the PF before enabling the VFs, so they require a PF driver to
> do that.
>
> Other SR-IOV devices don't require any PF setup before enabling VFs. Add a
> pci_sriov_configure_simple() interface so PF drivers for such devices can
> use it without repeating the VF-enabling code.
>
> Tested-by: Mark Rustad <mark.d.rustad@intel.com>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
> [bhelgaas: changelog, comment]
> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
> Reviewed-by: Greg Rose <gvrose8192@gmail.com>
> Reviewed-by: Christoph Hellwig <hch@lst.de>:wq
>
> commit a8ccf8a66663
> Author: Alexander Duyck <alexander.h.duyck@intel.com>
> Date: Tue Apr 24 16:47:16 2018 -0500
>
> PCI/IOV: Add pci-pf-stub driver for PFs that only enable VFs
>
> Some SR-IOV PF devices provide no functionality other than acting as a
> means of enabling VFs. For these devices, we want to enable the VFs and
> assign them to guest virtual machines, but there's no need to have a driver
> for the PF itself.
>
> Add a new pci-pf-stub driver to claim those PF devices and provide the
> generic VF enable functionality. An administrator can use the sysfs
> "sriov_numvfs" file to enable VFs, then assign them to guests.
>
> For now I only have one example ID provided by Amazon in terms of devices
> that require this functionality. The general idea is that in the future we
> will see other devices added as vendors come up with devices where the PF
> is more or less just a lightweight shim used to allocate VFs.
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
> [bhelgaas: changelog]
> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
> Reviewed-by: Greg Rose <gvrose8192@gmail.com>
> Reviewed-by: Christoph Hellwig <hch@lst.de>
>
> commit 115ddc491922
> Author: Alexander Duyck <alexander.h.duyck@intel.com>
> Date: Tue Apr 24 16:47:22 2018 -0500
>
> net: ena: Use pci_sriov_configure_simple() to enable VFs
>
> Instead of implementing our own version of a SR-IOV configuration stub in
> the ena driver, use the existing pci_sriov_configure_simple() function.
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
> Reviewed-by: Greg Rose <gvrose8192@gmail.com>
> Reviewed-by: Christoph Hellwig <hch@lst.de>
>
> commit 74d986abc20b
> Author: Alexander Duyck <alexander.h.duyck@intel.com>
> Date: Tue Apr 24 16:47:27 2018 -0500
>
> nvme-pci: Use pci_sriov_configure_simple() to enable VFs
>
> Instead of implementing our own version of a SR-IOV configuration stub in
> the nvme driver, use the existing pci_sriov_configure_simple() function.
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
> Reviewed-by: Christoph Hellwig <hch@lst.de>
^ permalink raw reply
* Re: [PATCH net-next 2/2 v2] netns: isolate seqnums to use per-netns locks
From: Christian Brauner @ 2018-04-24 22:20 UTC (permalink / raw)
To: Eric W. Biederman
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <87k1sw46i3.fsf@xmission.com>
On Tue, Apr 24, 2018 at 04:52:20PM -0500, Eric W. Biederman wrote:
> Christian Brauner <christian.brauner@ubuntu.com> writes:
>
> > Now that it's possible to have a different set of uevents in different
> > network namespaces, per-network namespace uevent sequence numbers are
> > introduced. This increases performance as locking is now restricted to the
> > network namespace affected by the uevent rather than locking everything.
> > Testing revealed significant performance improvements. For details see
> > "Testing" below.
>
> Maybe. Your locking is wrong, and a few other things are wrong. see
> below.
Thanks for the review! Happy to rework this until it's in a mergeable shape.
>
> > Since commit 692ec06 ("netns: send uevent messages") network namespaces not
> > owned by the intial user namespace can be sent uevents from a sufficiently
> > privileged userspace process.
> > In order to send a uevent into a network namespace not owned by the initial
> > user namespace we currently still need to take the *global mutex* that
> > locks the uevent socket list even though the list *only contains network
> > namespaces owned by the initial user namespace*. This needs to be done
> > because the uevent counter is a global variable. Taking the global lock is
> > performance sensitive since a user on the host can spawn a pool of n
> > process that each create their own new user and network namespaces and then
> > go on to inject uevents in parallel into the network namespace of all of
> > these processes. This can have a significant performance impact for the
> > host's udevd since it means that there can be a lot of delay between a
> > device being added and the corresponding uevent being sent out and
> > available for processing by udevd. It also means that each network
> > namespace not owned by the initial user namespace which userspace has sent
> > a uevent to will need to wait until the lock becomes available.
> >
> > Implementation:
> > This patch gives each network namespace its own uevent sequence number.
> > Each network namespace not owned by the initial user namespace receives its
> > own mutex. The struct uevent_sock is opaque to callers outside of kobject.c
> > so the mutex *can* and *is* only ever accessed in lib/kobject.c. In this
> > file it is clearly documented which lock has to be taken. All network
> > namespaces owned by the initial user namespace will still share the same
> > lock since they are all served sequentially via the uevent socket list.
> > This decouples the locking and ensures that the host retrieves uevents as
> > fast as possible even if there are a lot of uevents injected into network
> > namespaces not owned by the initial user namespace. In addition, each
> > network namespace not owned by the initial user namespace does not have to
> > wait on any other network namespace not sharing the same user namespace.
> >
> > Testing:
> > Two 4.17-rc1 test kernels were compiled. One with per netns uevent seqnums
> > with decoupled locking and one without. To ensure that testing made sense
> > both kernels carried the patch to remove network namespaces not owned by
> > the initial user namespace from the uevent socket list.
> > Three tests were constructed. All of them showed significant performance
> > improvements with per-netns uevent sequence numbers and decoupled locking.
> >
> > # Testcase 1:
> > Only Injecting Uevents into network namespaces not owned by the initial
> > user namespace.
> > - created 1000 new user namespace + network namespace pairs
> > - opened a uevent listener in each of those namespace pairs
> > - injected uevents into each of those network namespaces 10,000 times
> > meaning 10,000,000 (10 million) uevents were injected. (The high
> > number of uevent injections should get rid of a lot of jitter.)
> > The injection was done by fork()ing 1000 uevent injectors in a simple
> > for-loop to ensure that uevents were injected in parallel.
> > - mean transaction time was calculated:
> > - *without* uevent sequence number namespacing: 67 μs
> > - *with* uevent sequence number namespacing: 55 μs
> > - makes a difference of: 12 μs
> > - a t-test was performed on the two data vectors which revealed
> > shows significant performance improvements:
> > Welch Two Sample t-test
> > data: x1 and y1
> > t = 405.16, df = 18883000, p-value < 2.2e-16
> > alternative hypothesis: true difference in means is not equal to 0
> > 95 percent confidence interval:
> > 12.14949 12.26761
> > sample estimates:
> > mean of x mean of y
> > 68.48594 56.27739
> >
> > # Testcase 2:
> > Injecting Uevents into network namespaces not owned by the initial user
> > namespace and network namespaces owned by the initial user namespace.
> > - created 500 new user namespace + network namespace pairs
> > - created 500 new network namespace pairs
> > - opened a uevent listener in each of those namespace pairs
> > - injected uevents into each of those network namespaces 10,000 times
> > meaning 10,000,000 (10 million) uevents were injected. (The high
> > number of uevent injections should get rid of a lot of jitter.)
> > The injection was done by fork()ing 1000 uevent injectors in a simple
> > for-loop to ensure that uevents were injected in parallel.
> > - mean transaction time was calculated:
> > - *without* uevent sequence number namespacing: 572 μs
> > - *with* uevent sequence number namespacing: 514 μs
> > - makes a difference of: 58 μs
> > - a t-test was performed on the two data vectors which revealed
> > shows significant performance improvements:
> > Welch Two Sample t-test
> > data: x2 and y2
> > t = 38.685, df = 19682000, p-value < 2.2e-16
> > alternative hypothesis: true difference in means is not equal to 0
> > 95 percent confidence interval:
> > 55.10630 60.98815
> > sample estimates:
> > mean of x mean of y
> > 572.9684 514.9211
> >
> > # Testcase 3:
> > Created 500 new user namespace + network namespace pairs *without uevent
> > listeners*
> > - created 500 new network namespace pairs *without uevent listeners*
> > - injected uevents into each of those network namespaces 10,000 times
> > meaning 10,000,000 (10 million) uevents were injected. (The high number
> > of uevent injections should get rid of a lot of jitter.)
> > The injection was done by fork()ing 1000 uevent injectors in a simple
> > for-loop to ensure that uevents were injected in parallel.
> > - mean transaction time was calculated:
> > - *without* uevent sequence number namespacing: 206 μs
> > - *with* uevent sequence number namespacing: 163 μs
> > - makes a difference of: 43 μs
> > - a t-test was performed on the two data vectors which revealed
> > shows significant performance improvements:
> > Welch Two Sample t-test
> > data: x3 and y3
> > t = 58.37, df = 17711000, p-value < 2.2e-16
> > alternative hypothesis: true difference in means is not equal to 0
> > 95 percent confidence interval:
> > 41.77860 44.68178
> > sample estimates:
> > mean of x mean of y
> > 207.2632 164.0330
> >
> > Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> > ---
> > Changelog v1->v2:
> > * non-functional change: fix indendation for C directives in
> > kernel/ksysfs.c
> > Changelog v0->v1:
> > * add detailed test results to the commit message
> > * account for kernels compiled without CONFIG_NET
> > ---
> > include/linux/kobject.h | 2 +
> > include/net/net_namespace.h | 3 ++
> > kernel/ksysfs.c | 11 +++-
> > lib/kobject_uevent.c | 104 +++++++++++++++++++++++++++++-------
> > net/core/net_namespace.c | 14 +++++
> > 5 files changed, 114 insertions(+), 20 deletions(-)
> >
> > diff --git a/include/linux/kobject.h b/include/linux/kobject.h
> > index 7f6f93c3df9c..4e608968907f 100644
> > --- a/include/linux/kobject.h
> > +++ b/include/linux/kobject.h
> > @@ -36,8 +36,10 @@
> > extern char uevent_helper[];
> > #endif
> >
> > +#ifndef CONFIG_NET
> > /* counter to tag the uevent, read only except for the kobject core */
> > extern u64 uevent_seqnum;
> > +#endif
>
> That smells like an implementation bug somewhere.
Sorry, I'm not following. I'm I'm not mistaken there won't be any struct
net when CONFIG_NET=n. This has been reported by kbuild robot with alpha
and CONFIG_NET=n.
>
> > /*
> > * The actions here must match the index to the string array
> > diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
> > index 47e35cce3b64..e4e171b1ba69 100644
> > --- a/include/net/net_namespace.h
> > +++ b/include/net/net_namespace.h
> > @@ -85,6 +85,8 @@ struct net {
> > struct sock *genl_sock;
> >
> > struct uevent_sock *uevent_sock; /* uevent socket */
> > + /* counter to tag the uevent, read only except for the kobject core */
> > + u64 uevent_seqnum;
> >
> > struct list_head dev_base_head;
> > struct hlist_head *dev_name_head;
> > @@ -189,6 +191,7 @@ extern struct list_head net_namespace_list;
> >
> > struct net *get_net_ns_by_pid(pid_t pid);
> > struct net *get_net_ns_by_fd(int fd);
> > +u64 get_ns_uevent_seqnum_by_vpid(void);
> >
> > #ifdef CONFIG_SYSCTL
> > void ipx_register_sysctl(void);
> > diff --git a/kernel/ksysfs.c b/kernel/ksysfs.c
> > index 46ba853656f6..38b70b90a21f 100644
> > --- a/kernel/ksysfs.c
> > +++ b/kernel/ksysfs.c
> > @@ -19,6 +19,7 @@
> > #include <linux/sched.h>
> > #include <linux/capability.h>
> > #include <linux/compiler.h>
> > +#include <net/net_namespace.h>
> >
> > #include <linux/rcupdate.h> /* rcu_expedited and rcu_normal */
> >
> > @@ -33,7 +34,15 @@ static struct kobj_attribute _name##_attr = \
> > static ssize_t uevent_seqnum_show(struct kobject *kobj,
> > struct kobj_attribute *attr, char *buf)
> > {
> > - return sprintf(buf, "%llu\n", (unsigned long long)uevent_seqnum);
> > + u64 seqnum;
> > +
> > +#ifdef CONFIG_NET
> > + seqnum = get_ns_uevent_seqnum_by_vpid();
> > +#else
> > + seqnum = uevent_seqnum;
> > +#endif
>
> This can be simplified to be just:
> seqnum = current->nsproxy->net_ns->uevent_seqnum;
Does that work even if CONFIG_NET=n?
>
> Except that is not correct either. As every instance of sysfs has a
> network namespace associated with it, and you are not fetching that
> network namespace.
I'm not yet familiar with all aspects of sysfs so thanks for pointing
that out. Then I'll try to come up with a way to fetch the network
namespace associated with sysfs. Unless you already know exactly how to
do this and can point it out.
This would also lets us drop get_ns_uevent_seqnum_by_vpid().
>
> Typically this would call for making this file per network namespace
> so you would have this information available. Sigh. I don't know if
> there is an easy way to do that for this file.
>
> > +
> > + return sprintf(buf, "%llu\n", (unsigned long long)seqnum);
> > }
> > KERNEL_ATTR_RO(uevent_seqnum);
> >
> > diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> > index f5f5038787ac..5da20def556d 100644
> > --- a/lib/kobject_uevent.c
> > +++ b/lib/kobject_uevent.c
> > @@ -29,21 +29,42 @@
> > #include <net/net_namespace.h>
> >
> >
> > +#ifndef CONFIG_NET
> > u64 uevent_seqnum;
> > +#endif
> > +
> > #ifdef CONFIG_UEVENT_HELPER
> > char uevent_helper[UEVENT_HELPER_PATH_LEN] = CONFIG_UEVENT_HELPER_PATH;
> > #endif
> >
> > +/*
> > + * Size a buffer needs to be in order to hold the largest possible sequence
> > + * number stored in a u64 including \0 byte: 2^64 - 1 = 21 chars.
> > + */
> > +#define SEQNUM_BUFSIZE (sizeof("SEQNUM=") + 21)
> > struct uevent_sock {
> > struct list_head list;
> > struct sock *sk;
> > + /*
> > + * This mutex protects uevent sockets and the uevent counter of
> > + * network namespaces *not* owned by init_user_ns.
> > + * For network namespaces owned by init_user_ns this lock is *not*
> > + * valid instead the global uevent_sock_mutex must be used!
> > + */
> > + struct mutex sk_mutex;
> > };
> >
> > #ifdef CONFIG_NET
> > static LIST_HEAD(uevent_sock_list);
> > #endif
> >
> > -/* This lock protects uevent_seqnum and uevent_sock_list */
> > +/*
> > + * This mutex protects uevent sockets and the uevent counter of network
> > + * namespaces owned by init_user_ns.
> > + * For network namespaces not owned by init_user_ns this lock is *not*
> > + * valid instead the network namespace specific sk_mutex in struct
> > + * uevent_sock must be used!
> > + */
> > static DEFINE_MUTEX(uevent_sock_mutex);
> >
> > /* the strings here must match the enum in include/linux/kobject.h */
> > @@ -253,6 +274,22 @@ static int kobj_bcast_filter(struct sock *dsk, struct sk_buff *skb, void *data)
> >
> > return 0;
> > }
> > +
> > +static bool can_hold_seqnum(const struct kobj_uevent_env *env, size_t len)
> > +{
> > + if (env->envp_idx >= ARRAY_SIZE(env->envp)) {
> > + WARN(1, KERN_ERR "Failed to append sequence number. "
> > + "Too many uevent variables\n");
> > + return false;
> > + }
> > +
> > + if ((env->buflen + len) > UEVENT_BUFFER_SIZE) {
> > + WARN(1, KERN_ERR "Insufficient space to append sequence number\n");
> > + return false;
> > + }
> > +
> > + return true;
> > +}
> > #endif
> >
> > #ifdef CONFIG_UEVENT_HELPER
> > @@ -308,18 +345,22 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
> >
> > /* send netlink message */
> > list_for_each_entry(ue_sk, &uevent_sock_list, list) {
> > + /* bump sequence number */
> > + u64 seqnum = ++sock_net(ue_sk->sk)->uevent_seqnum;
> > struct sock *uevent_sock = ue_sk->sk;
> > + char buf[SEQNUM_BUFSIZE];
> >
> > if (!netlink_has_listeners(uevent_sock, 1))
> > continue;
> >
> > if (!skb) {
> > - /* allocate message with the maximum possible size */
> > + /* calculate header length */
> > size_t len = strlen(action_string) + strlen(devpath) + 2;
> > char *scratch;
> >
> > + /* allocate message with the maximum possible size */
> > retval = -ENOMEM;
> > - skb = alloc_skb(len + env->buflen, GFP_KERNEL);
> > + skb = alloc_skb(len + env->buflen + SEQNUM_BUFSIZE, GFP_KERNEL);
> > if (!skb)
> > continue;
> >
> > @@ -327,11 +368,24 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
> > scratch = skb_put(skb, len);
> > sprintf(scratch, "%s@%s", action_string, devpath);
> >
> > + /* add env */
> > skb_put_data(skb, env->buf, env->buflen);
> >
> > NETLINK_CB(skb).dst_group = 1;
> > }
> >
> > + /* prepare netns seqnum */
> > + retval = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu", seqnum);
> > + if (retval < 0 || retval >= SEQNUM_BUFSIZE)
> > + continue;
> > + retval++;
> > +
> > + if (!can_hold_seqnum(env, retval))
> > + continue;
>
> You have allocated enough space in the skb why does can_hold_seqnum make
> sense?
Because it doesn't check whether the socket buffer can hold the sequence
number but checks whether the uevent buffer size in "env" can hold it.
uevents are only delivered if the env buffer is large enough to hold all
of the info including the sequence number. That's independent of the
socket buffer.
>
> Do you need to back seqnum out of the env later for this to work twice
> in a row?
I guess I can just override it. It just felt cleaner to trim it.
>
> > +
> > + /* append netns seqnum */
> > + skb_put_data(skb, buf, retval);
> > +
> > retval = netlink_broadcast_filtered(uevent_sock, skb_get(skb),
> > 0, 1, GFP_KERNEL,
> > kobj_bcast_filter,
> > @@ -339,8 +393,13 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
> > /* ENOBUFS should be handled in userspace */
> > if (retval == -ENOBUFS || retval == -ESRCH)
> > retval = 0;
> > +
> > + /* remove netns seqnum */
> > + skb_trim(skb, env->buflen);
>
> Have you checked to see if the seqnum actually makes it to userspace.
Yes, I did. I also wonder why it wouldn't make it. Any specific reason
why you suspect this?
> > }
> > consume_skb(skb);
> > +#else
> > + uevent_seqnum++;
> > #endif
> > return retval;
> > }
> > @@ -510,14 +569,7 @@ int kobject_uevent_env(struct kobject *kobj, enum kobject_action action,
> > }
> >
> > mutex_lock(&uevent_sock_mutex);
> > - /* we will send an event, so request a new sequence number */
> > - retval = add_uevent_var(env, "SEQNUM=%llu", (unsigned long long)++uevent_seqnum);
> > - if (retval) {
> > - mutex_unlock(&uevent_sock_mutex);
> > - goto exit;
> > - }
> > - retval = kobject_uevent_net_broadcast(kobj, env, action_string,
> > - devpath);
> > + retval = kobject_uevent_net_broadcast(kobj, env, action_string, devpath);
> > mutex_unlock(&uevent_sock_mutex);
>
> How does all of this work with events for network devices that are not
> in the initial network namespace. This looks to me like this code fails
> to take the sk_mutex.
But in this list only non-initial network namespaces that are owned by
the initial user namespace are recorded and for these uevent_sock_mutex
has to be taken. Am I missing something?
>
> > #ifdef CONFIG_UEVENT_HELPER
> > @@ -605,17 +657,18 @@ int add_uevent_var(struct kobj_uevent_env *env, const char *format, ...)
> > EXPORT_SYMBOL_GPL(add_uevent_var);
> >
> > #if defined(CONFIG_NET)
> > -static int uevent_net_broadcast(struct sock *usk, struct sk_buff *skb,
> > +static int uevent_net_broadcast(struct uevent_sock *ue_sk, struct sk_buff *skb,
> > struct netlink_ext_ack *extack)
> > {
> > - /* u64 to chars: 2^64 - 1 = 21 chars */
> > - char buf[sizeof("SEQNUM=") + 21];
> > + struct sock *usk = ue_sk->sk;
> > + char buf[SEQNUM_BUFSIZE];
> > struct sk_buff *skbc;
> > int ret;
> >
> > /* bump and prepare sequence number */
> > - ret = snprintf(buf, sizeof(buf), "SEQNUM=%llu", ++uevent_seqnum);
> > - if (ret < 0 || (size_t)ret >= sizeof(buf))
> > + ret = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu",
> > + ++sock_net(ue_sk->sk)->uevent_seqnum);
> > + if (ret < 0 || ret >= SEQNUM_BUFSIZE)
> > return -ENOMEM;
> > ret++;
> >
> > @@ -668,9 +721,15 @@ static int uevent_net_rcv_skb(struct sk_buff *skb, struct nlmsghdr *nlh,
> > return -EPERM;
> > }
> >
> > - mutex_lock(&uevent_sock_mutex);
> > - ret = uevent_net_broadcast(net->uevent_sock->sk, skb, extack);
> > - mutex_unlock(&uevent_sock_mutex);
> > + if (net->user_ns == &init_user_ns)
> > + mutex_lock(&uevent_sock_mutex);
> > + else
> > + mutex_lock(&net->uevent_sock->sk_mutex);
> > + ret = uevent_net_broadcast(net->uevent_sock, skb, extack);
> > + if (net->user_ns == &init_user_ns)
> > + mutex_unlock(&uevent_sock_mutex);
> > + else
> > + mutex_unlock(&net->uevent_sock->sk_mutex);
> >
> > return ret;
> > }
> > @@ -708,6 +767,13 @@ static int uevent_net_init(struct net *net)
> > mutex_lock(&uevent_sock_mutex);
> > list_add_tail(&ue_sk->list, &uevent_sock_list);
> > mutex_unlock(&uevent_sock_mutex);
> > + } else {
> > + /*
> > + * Uevent sockets and counters for network namespaces
> > + * not owned by the initial user namespace have their
> > + * own mutex.
> > + */
> > + mutex_init(&ue_sk->sk_mutex);
> > }
> >
> > return 0;
> > diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
> > index a11e03f920d3..8894638f5150 100644
> > --- a/net/core/net_namespace.c
> > +++ b/net/core/net_namespace.c
> > @@ -618,6 +618,20 @@ struct net *get_net_ns_by_pid(pid_t pid)
> > }
> > EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
> >
> > +u64 get_ns_uevent_seqnum_by_vpid(void)
> > +{
> > + pid_t cur_pid;
> > + struct net *net;
> > +
> > + cur_pid = task_pid_vnr(current);
> > + net = get_net_ns_by_pid(cur_pid);
> > + if (IS_ERR(net))
> > + return 0;
> > +
> > + return net->uevent_seqnum;
> > +}
> > +EXPORT_SYMBOL_GPL(get_ns_uevent_seqnum_by_vpid);
>
> I just have to say this function is completely crazy.
> You go from the tsk to the pid back to the tsk.
> And you leak a struct net pointer.
>
> It is much simpler and less racy to say:
>
> current->nsproxy->net_ns->uevent_seqnum;
>
> That you are accessing current->nsproxy means nsproxy can't
> change. The rcu_read_lock etc that get_net_ns_by_pid does
> is there for accessing non-current tasks.
>
>
>
> > static __net_init int net_ns_net_init(struct net *net)
> > {
> > #ifdef CONFIG_NET_NS
^ permalink raw reply
* [PATCH] selftests: net: add in_netns.sh TEST_GEN_PROGS_EXTENDED
From: Anders Roxell @ 2018-04-24 22:19 UTC (permalink / raw)
To: davem, shuah; +Cc: netdev, linux-kselftest, linux-kernel, Anders Roxell
Script in_netns.sh is a utility function and not its own test so it
shouldn't be part of the TEST_PROGS. The in_netns.sh get used by
run_afpackettests.
To install in_netns.sh without being added to the main run_kselftest.sh
script use the TEST_GEN_PROGS_EXTENDED variable.
Fixes: 5ff9c1a3dd92 ("selftests: net: add in_netns.sh to TEST_PROGS")
Signed-off-by: Anders Roxell <anders.roxell@linaro.org>
---
tools/testing/selftests/net/Makefile | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/net/Makefile b/tools/testing/selftests/net/Makefile
index c3761c35f542..4a8cfe8071a7 100644
--- a/tools/testing/selftests/net/Makefile
+++ b/tools/testing/selftests/net/Makefile
@@ -5,7 +5,8 @@ CFLAGS = -Wall -Wl,--no-as-needed -O2 -g
CFLAGS += -I../../../../usr/include/
TEST_PROGS := run_netsocktests run_afpackettests test_bpf.sh netdevice.sh rtnetlink.sh
-TEST_PROGS += fib_tests.sh fib-onlink-tests.sh in_netns.sh pmtu.sh
+TEST_PROGS += fib_tests.sh fib-onlink-tests.sh pmtu.sh
+TEST_GEN_PROGS_EXTENDED := in_netns.sh
TEST_GEN_FILES = socket
TEST_GEN_FILES += psock_fanout psock_tpacket msg_zerocopy
TEST_GEN_FILES += tcp_mmap
--
2.11.0
^ permalink raw reply related
* Re: [PATCH bpf-next] tools/bpf: remove test_sock_addr from TEST_GEN_PROGS
From: Daniel Borkmann @ 2018-04-24 22:01 UTC (permalink / raw)
To: Yonghong Song, ast, netdev; +Cc: kernel-team
In-Reply-To: <20180424214504.3000168-1-yhs@fb.com>
On 04/24/2018 11:45 PM, Yonghong Song wrote:
> Since test_sock_addr is not supposed to run by itself,
> remove it from TEST_GEN_PROGS and add it to
> TEST_GEN_PROGS_EXTENDED. This way, run_tests will
> not run test_sock_addr. The corresponding test to run
> is test_sock_addr.sh.
>
> Signed-off-by: Yonghong Song <yhs@fb.com>
Applied to bpf-next, thanks Yonghong!
^ permalink raw reply
* Re: [PATCH v2] selftests: bpf: update .gitignore with missing file
From: Daniel Borkmann @ 2018-04-24 21:56 UTC (permalink / raw)
To: Anders Roxell, ast, shuah; +Cc: netdev, linux-kernel, linux-kselftest
In-Reply-To: <20180423225305.18947-1-anders.roxell@linaro.org>
On 04/24/2018 12:53 AM, Anders Roxell wrote:
> Fixes: c0fa1b6c3efc ("bpf: btf: Add BTF tests")
> Signed-off-by: Anders Roxell <anders.roxell@linaro.org>
> ---
> Rebased against bpf-next.
Applied to bpf-next, thanks Anders!
^ permalink raw reply
* Re: [PATCH net-next 1/2 v2] netns: restrict uevents
From: Eric W. Biederman @ 2018-04-24 21:54 UTC (permalink / raw)
To: Christian Brauner
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <20180424204335.12904-2-christian.brauner@ubuntu.com>
We already do this in practice in userspace. It doesn't make much
sense to perform this delivery. So we might as well make this optimization.
Christian Brauner <christian.brauner@ubuntu.com> writes:
> commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
>
> enabled sending hotplug events into all network namespaces back in 2010.
> Over time the set of uevents that get sent into all network namespaces has
> shrunk a little. We have now reached the point where hotplug events for all
> devices that carry a namespace tag are filtered according to that
> namespace. Specifically, they are filtered whenever the namespace tag of
> the kobject does not match the namespace tag of the netlink socket. One
> example are network devices. Uevents for network devices only show up in
> the network namespaces these devices are moved to or created in.
>
> However, any uevent for a kobject that does not have a namespace tag
> associated with it will not be filtered and we will broadcast it into all
> network namespaces. This behavior stopped making sense when user namespaces
> were introduced.
>
> This patch restricts uevents to the initial user namespace for a couple of
> reasons that have been extensively discusses on the mailing list [1].
> - Thundering herd:
> Broadcasting uevents into all network namespaces introduces significant
> overhead.
> All processes that listen to uevents running in non-initial user
> namespaces will end up responding to uevents that will be meaningless to
> them. Mainly, because non-initial user namespaces cannot easily manage
> devices unless they have a privileged host-process helping them out. This
> means that there will be a thundering herd of activity when there
> shouldn't be any.
> - Uevents from non-root users are already filtered in userspace:
> Uevents are filtered by userspace in a user namespace because the
> received uid != 0. Instead the uid associated with the event will be
> 65534 == "nobody" because the global root uid is not mapped.
> This means we can safely and without introducing regressions modify the
> kernel to not send uevents into all network namespaces whose owning user
> namespace is not the initial user namespace because we know that
> userspace will ignore the message because of the uid anyway. I have
> a) verified that is is true for every udev implementation out there b)
> that this behavior has been present in all udev implementations from the
> very beginning.
> - Removing needless overhead/Increasing performance:
> Currently, the uevent socket for each network namespace is added to the
> global variable uevent_sock_list. The list itself needs to be protected
> by a mutex. So everytime a uevent is generated the mutex is taken on the
> list. The mutex is held *from the creation of the uevent (memory
> allocation, string creation etc. until all uevent sockets have been
> handled*. This is aggravated by the fact that for each uevent socket that
> has listeners the mc_list must be walked as well which means we're
> talking O(n^2) here. Given that a standard Linux workload usually has
> quite a lot of network namespaces and - in the face of containers - a lot
> of user namespaces this quickly becomes a performance problem (see
> "Thundering herd" above). By just recording uevent sockets of network
> namespaces that are owned by the initial user namespace we significantly
> increase performance in this codepath.
> - Injecting uevents:
> There's a valid argument that containers might be interested in receiving
> device events especially if they are delegated to them by a privileged
> userspace process. One prime example are SR-IOV enabled devices that are
> explicitly designed to be handed of to other users such as VMs or
> containers.
> This use-case can now be correctly handled since
> commit 692ec06d7c92 ("netns: send uevent messages"). This commit
> introduced the ability to send uevents from userspace. As such we can let
> a sufficiently privileged (CAP_SYS_ADMIN in the owning user namespace of
> the network namespace of the netlink socket) userspace process make a
> decision what uevents should be sent. This removes the need to blindly
> broadcast uevents into all user namespaces and provides a performant and
> safe solution to this problem.
> - Filtering logic:
> This patch filters by *owning user namespace of the network namespace a
> given task resides in* and not by user namespace of the task per se. This
> means if the user namespace of a given task is unshared but the network
> namespace is kept and is owned by the initial user namespace a listener
> that is opening the uevent socket in that network namespace can still
> listen to uevents.
>
> [1]: https://lkml.org/lkml/2018/4/4/739
> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Acked-by: "Eric W. Biederman" <ebiederm@xmission.com>
> ---
> Changelog v1->v2:
> * patch unchanged
> Changelog v0->v1:
> * patch unchanged
> ---
> lib/kobject_uevent.c | 18 ++++++++++++------
> 1 file changed, 12 insertions(+), 6 deletions(-)
>
> diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> index 15ea216a67ce..f5f5038787ac 100644
> --- a/lib/kobject_uevent.c
> +++ b/lib/kobject_uevent.c
> @@ -703,9 +703,13 @@ static int uevent_net_init(struct net *net)
>
> net->uevent_sock = ue_sk;
>
> - mutex_lock(&uevent_sock_mutex);
> - list_add_tail(&ue_sk->list, &uevent_sock_list);
> - mutex_unlock(&uevent_sock_mutex);
> + /* Restrict uevents to initial user namespace. */
> + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> + mutex_lock(&uevent_sock_mutex);
> + list_add_tail(&ue_sk->list, &uevent_sock_list);
> + mutex_unlock(&uevent_sock_mutex);
> + }
> +
> return 0;
> }
>
> @@ -713,9 +717,11 @@ static void uevent_net_exit(struct net *net)
> {
> struct uevent_sock *ue_sk = net->uevent_sock;
>
> - mutex_lock(&uevent_sock_mutex);
> - list_del(&ue_sk->list);
> - mutex_unlock(&uevent_sock_mutex);
> + if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
> + mutex_lock(&uevent_sock_mutex);
> + list_del(&ue_sk->list);
> + mutex_unlock(&uevent_sock_mutex);
> + }
>
> netlink_kernel_release(ue_sk->sk);
> kfree(ue_sk);
^ permalink raw reply
* Re: [PATCH bpf-next v5 0/2] bpf: allow map helpers access to map values directly
From: Daniel Borkmann @ 2018-04-24 21:54 UTC (permalink / raw)
To: Paul Chaignon, Alexei Starovoitov, netdev; +Cc: iovisor-dev, paul.chaignon
In-Reply-To: <20180424130636.GA6270@Nover>
On 04/24/2018 03:06 PM, Paul Chaignon wrote:
> Currently, helpers that expect ARG_PTR_TO_MAP_KEY and ARG_PTR_TO_MAP_VALUE
> can only access stack and packet memory. This patchset allows these
> helpers to directly access map values by passing registers of type
> PTR_TO_MAP_VALUE.
>
> The first patch changes the verifier; the second adds new test cases.
>
> The first three versions of this patchset were sent on the iovisor-dev
> mailing list only.
>
> Changelogs:
> Changes in v5:
> - Refactor using check_helper_mem_access.
> Changes in v4:
> - Rebase.
> Changes in v3:
> - Bug fixes.
> - Negative test cases.
> Changes in v2:
> - Additional test cases for adjusted maps.
Applied to bpf-next, thanks Paul!
^ permalink raw reply
* Re: [PATCH net-next 2/2 v2] netns: isolate seqnums to use per-netns locks
From: Eric W. Biederman @ 2018-04-24 21:52 UTC (permalink / raw)
To: Christian Brauner
Cc: davem, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <20180424204335.12904-3-christian.brauner@ubuntu.com>
Christian Brauner <christian.brauner@ubuntu.com> writes:
> Now that it's possible to have a different set of uevents in different
> network namespaces, per-network namespace uevent sequence numbers are
> introduced. This increases performance as locking is now restricted to the
> network namespace affected by the uevent rather than locking everything.
> Testing revealed significant performance improvements. For details see
> "Testing" below.
Maybe. Your locking is wrong, and a few other things are wrong. see
below.
> Since commit 692ec06 ("netns: send uevent messages") network namespaces not
> owned by the intial user namespace can be sent uevents from a sufficiently
> privileged userspace process.
> In order to send a uevent into a network namespace not owned by the initial
> user namespace we currently still need to take the *global mutex* that
> locks the uevent socket list even though the list *only contains network
> namespaces owned by the initial user namespace*. This needs to be done
> because the uevent counter is a global variable. Taking the global lock is
> performance sensitive since a user on the host can spawn a pool of n
> process that each create their own new user and network namespaces and then
> go on to inject uevents in parallel into the network namespace of all of
> these processes. This can have a significant performance impact for the
> host's udevd since it means that there can be a lot of delay between a
> device being added and the corresponding uevent being sent out and
> available for processing by udevd. It also means that each network
> namespace not owned by the initial user namespace which userspace has sent
> a uevent to will need to wait until the lock becomes available.
>
> Implementation:
> This patch gives each network namespace its own uevent sequence number.
> Each network namespace not owned by the initial user namespace receives its
> own mutex. The struct uevent_sock is opaque to callers outside of kobject.c
> so the mutex *can* and *is* only ever accessed in lib/kobject.c. In this
> file it is clearly documented which lock has to be taken. All network
> namespaces owned by the initial user namespace will still share the same
> lock since they are all served sequentially via the uevent socket list.
> This decouples the locking and ensures that the host retrieves uevents as
> fast as possible even if there are a lot of uevents injected into network
> namespaces not owned by the initial user namespace. In addition, each
> network namespace not owned by the initial user namespace does not have to
> wait on any other network namespace not sharing the same user namespace.
>
> Testing:
> Two 4.17-rc1 test kernels were compiled. One with per netns uevent seqnums
> with decoupled locking and one without. To ensure that testing made sense
> both kernels carried the patch to remove network namespaces not owned by
> the initial user namespace from the uevent socket list.
> Three tests were constructed. All of them showed significant performance
> improvements with per-netns uevent sequence numbers and decoupled locking.
>
> # Testcase 1:
> Only Injecting Uevents into network namespaces not owned by the initial
> user namespace.
> - created 1000 new user namespace + network namespace pairs
> - opened a uevent listener in each of those namespace pairs
> - injected uevents into each of those network namespaces 10,000 times
> meaning 10,000,000 (10 million) uevents were injected. (The high
> number of uevent injections should get rid of a lot of jitter.)
> The injection was done by fork()ing 1000 uevent injectors in a simple
> for-loop to ensure that uevents were injected in parallel.
> - mean transaction time was calculated:
> - *without* uevent sequence number namespacing: 67 μs
> - *with* uevent sequence number namespacing: 55 μs
> - makes a difference of: 12 μs
> - a t-test was performed on the two data vectors which revealed
> shows significant performance improvements:
> Welch Two Sample t-test
> data: x1 and y1
> t = 405.16, df = 18883000, p-value < 2.2e-16
> alternative hypothesis: true difference in means is not equal to 0
> 95 percent confidence interval:
> 12.14949 12.26761
> sample estimates:
> mean of x mean of y
> 68.48594 56.27739
>
> # Testcase 2:
> Injecting Uevents into network namespaces not owned by the initial user
> namespace and network namespaces owned by the initial user namespace.
> - created 500 new user namespace + network namespace pairs
> - created 500 new network namespace pairs
> - opened a uevent listener in each of those namespace pairs
> - injected uevents into each of those network namespaces 10,000 times
> meaning 10,000,000 (10 million) uevents were injected. (The high
> number of uevent injections should get rid of a lot of jitter.)
> The injection was done by fork()ing 1000 uevent injectors in a simple
> for-loop to ensure that uevents were injected in parallel.
> - mean transaction time was calculated:
> - *without* uevent sequence number namespacing: 572 μs
> - *with* uevent sequence number namespacing: 514 μs
> - makes a difference of: 58 μs
> - a t-test was performed on the two data vectors which revealed
> shows significant performance improvements:
> Welch Two Sample t-test
> data: x2 and y2
> t = 38.685, df = 19682000, p-value < 2.2e-16
> alternative hypothesis: true difference in means is not equal to 0
> 95 percent confidence interval:
> 55.10630 60.98815
> sample estimates:
> mean of x mean of y
> 572.9684 514.9211
>
> # Testcase 3:
> Created 500 new user namespace + network namespace pairs *without uevent
> listeners*
> - created 500 new network namespace pairs *without uevent listeners*
> - injected uevents into each of those network namespaces 10,000 times
> meaning 10,000,000 (10 million) uevents were injected. (The high number
> of uevent injections should get rid of a lot of jitter.)
> The injection was done by fork()ing 1000 uevent injectors in a simple
> for-loop to ensure that uevents were injected in parallel.
> - mean transaction time was calculated:
> - *without* uevent sequence number namespacing: 206 μs
> - *with* uevent sequence number namespacing: 163 μs
> - makes a difference of: 43 μs
> - a t-test was performed on the two data vectors which revealed
> shows significant performance improvements:
> Welch Two Sample t-test
> data: x3 and y3
> t = 58.37, df = 17711000, p-value < 2.2e-16
> alternative hypothesis: true difference in means is not equal to 0
> 95 percent confidence interval:
> 41.77860 44.68178
> sample estimates:
> mean of x mean of y
> 207.2632 164.0330
>
> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> ---
> Changelog v1->v2:
> * non-functional change: fix indendation for C directives in
> kernel/ksysfs.c
> Changelog v0->v1:
> * add detailed test results to the commit message
> * account for kernels compiled without CONFIG_NET
> ---
> include/linux/kobject.h | 2 +
> include/net/net_namespace.h | 3 ++
> kernel/ksysfs.c | 11 +++-
> lib/kobject_uevent.c | 104 +++++++++++++++++++++++++++++-------
> net/core/net_namespace.c | 14 +++++
> 5 files changed, 114 insertions(+), 20 deletions(-)
>
> diff --git a/include/linux/kobject.h b/include/linux/kobject.h
> index 7f6f93c3df9c..4e608968907f 100644
> --- a/include/linux/kobject.h
> +++ b/include/linux/kobject.h
> @@ -36,8 +36,10 @@
> extern char uevent_helper[];
> #endif
>
> +#ifndef CONFIG_NET
> /* counter to tag the uevent, read only except for the kobject core */
> extern u64 uevent_seqnum;
> +#endif
That smells like an implementation bug somewhere.
> /*
> * The actions here must match the index to the string array
> diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
> index 47e35cce3b64..e4e171b1ba69 100644
> --- a/include/net/net_namespace.h
> +++ b/include/net/net_namespace.h
> @@ -85,6 +85,8 @@ struct net {
> struct sock *genl_sock;
>
> struct uevent_sock *uevent_sock; /* uevent socket */
> + /* counter to tag the uevent, read only except for the kobject core */
> + u64 uevent_seqnum;
>
> struct list_head dev_base_head;
> struct hlist_head *dev_name_head;
> @@ -189,6 +191,7 @@ extern struct list_head net_namespace_list;
>
> struct net *get_net_ns_by_pid(pid_t pid);
> struct net *get_net_ns_by_fd(int fd);
> +u64 get_ns_uevent_seqnum_by_vpid(void);
>
> #ifdef CONFIG_SYSCTL
> void ipx_register_sysctl(void);
> diff --git a/kernel/ksysfs.c b/kernel/ksysfs.c
> index 46ba853656f6..38b70b90a21f 100644
> --- a/kernel/ksysfs.c
> +++ b/kernel/ksysfs.c
> @@ -19,6 +19,7 @@
> #include <linux/sched.h>
> #include <linux/capability.h>
> #include <linux/compiler.h>
> +#include <net/net_namespace.h>
>
> #include <linux/rcupdate.h> /* rcu_expedited and rcu_normal */
>
> @@ -33,7 +34,15 @@ static struct kobj_attribute _name##_attr = \
> static ssize_t uevent_seqnum_show(struct kobject *kobj,
> struct kobj_attribute *attr, char *buf)
> {
> - return sprintf(buf, "%llu\n", (unsigned long long)uevent_seqnum);
> + u64 seqnum;
> +
> +#ifdef CONFIG_NET
> + seqnum = get_ns_uevent_seqnum_by_vpid();
> +#else
> + seqnum = uevent_seqnum;
> +#endif
This can be simplified to be just:
seqnum = current->nsproxy->net_ns->uevent_seqnum;
Except that is not correct either. As every instance of sysfs has a
network namespace associated with it, and you are not fetching that
network namespace.
Typically this would call for making this file per network namespace
so you would have this information available. Sigh. I don't know if
there is an easy way to do that for this file.
> +
> + return sprintf(buf, "%llu\n", (unsigned long long)seqnum);
> }
> KERNEL_ATTR_RO(uevent_seqnum);
>
> diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
> index f5f5038787ac..5da20def556d 100644
> --- a/lib/kobject_uevent.c
> +++ b/lib/kobject_uevent.c
> @@ -29,21 +29,42 @@
> #include <net/net_namespace.h>
>
>
> +#ifndef CONFIG_NET
> u64 uevent_seqnum;
> +#endif
> +
> #ifdef CONFIG_UEVENT_HELPER
> char uevent_helper[UEVENT_HELPER_PATH_LEN] = CONFIG_UEVENT_HELPER_PATH;
> #endif
>
> +/*
> + * Size a buffer needs to be in order to hold the largest possible sequence
> + * number stored in a u64 including \0 byte: 2^64 - 1 = 21 chars.
> + */
> +#define SEQNUM_BUFSIZE (sizeof("SEQNUM=") + 21)
> struct uevent_sock {
> struct list_head list;
> struct sock *sk;
> + /*
> + * This mutex protects uevent sockets and the uevent counter of
> + * network namespaces *not* owned by init_user_ns.
> + * For network namespaces owned by init_user_ns this lock is *not*
> + * valid instead the global uevent_sock_mutex must be used!
> + */
> + struct mutex sk_mutex;
> };
>
> #ifdef CONFIG_NET
> static LIST_HEAD(uevent_sock_list);
> #endif
>
> -/* This lock protects uevent_seqnum and uevent_sock_list */
> +/*
> + * This mutex protects uevent sockets and the uevent counter of network
> + * namespaces owned by init_user_ns.
> + * For network namespaces not owned by init_user_ns this lock is *not*
> + * valid instead the network namespace specific sk_mutex in struct
> + * uevent_sock must be used!
> + */
> static DEFINE_MUTEX(uevent_sock_mutex);
>
> /* the strings here must match the enum in include/linux/kobject.h */
> @@ -253,6 +274,22 @@ static int kobj_bcast_filter(struct sock *dsk, struct sk_buff *skb, void *data)
>
> return 0;
> }
> +
> +static bool can_hold_seqnum(const struct kobj_uevent_env *env, size_t len)
> +{
> + if (env->envp_idx >= ARRAY_SIZE(env->envp)) {
> + WARN(1, KERN_ERR "Failed to append sequence number. "
> + "Too many uevent variables\n");
> + return false;
> + }
> +
> + if ((env->buflen + len) > UEVENT_BUFFER_SIZE) {
> + WARN(1, KERN_ERR "Insufficient space to append sequence number\n");
> + return false;
> + }
> +
> + return true;
> +}
> #endif
>
> #ifdef CONFIG_UEVENT_HELPER
> @@ -308,18 +345,22 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
>
> /* send netlink message */
> list_for_each_entry(ue_sk, &uevent_sock_list, list) {
> + /* bump sequence number */
> + u64 seqnum = ++sock_net(ue_sk->sk)->uevent_seqnum;
> struct sock *uevent_sock = ue_sk->sk;
> + char buf[SEQNUM_BUFSIZE];
>
> if (!netlink_has_listeners(uevent_sock, 1))
> continue;
>
> if (!skb) {
> - /* allocate message with the maximum possible size */
> + /* calculate header length */
> size_t len = strlen(action_string) + strlen(devpath) + 2;
> char *scratch;
>
> + /* allocate message with the maximum possible size */
> retval = -ENOMEM;
> - skb = alloc_skb(len + env->buflen, GFP_KERNEL);
> + skb = alloc_skb(len + env->buflen + SEQNUM_BUFSIZE, GFP_KERNEL);
> if (!skb)
> continue;
>
> @@ -327,11 +368,24 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
> scratch = skb_put(skb, len);
> sprintf(scratch, "%s@%s", action_string, devpath);
>
> + /* add env */
> skb_put_data(skb, env->buf, env->buflen);
>
> NETLINK_CB(skb).dst_group = 1;
> }
>
> + /* prepare netns seqnum */
> + retval = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu", seqnum);
> + if (retval < 0 || retval >= SEQNUM_BUFSIZE)
> + continue;
> + retval++;
> +
> + if (!can_hold_seqnum(env, retval))
> + continue;
You have allocated enough space in the skb why does can_hold_seqnum make
sense?
Do you need to back seqnum out of the env later for this to work twice
in a row?
> +
> + /* append netns seqnum */
> + skb_put_data(skb, buf, retval);
> +
> retval = netlink_broadcast_filtered(uevent_sock, skb_get(skb),
> 0, 1, GFP_KERNEL,
> kobj_bcast_filter,
> @@ -339,8 +393,13 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
> /* ENOBUFS should be handled in userspace */
> if (retval == -ENOBUFS || retval == -ESRCH)
> retval = 0;
> +
> + /* remove netns seqnum */
> + skb_trim(skb, env->buflen);
Have you checked to see if the seqnum actually makes it to userspace.
> }
> consume_skb(skb);
> +#else
> + uevent_seqnum++;
> #endif
> return retval;
> }
> @@ -510,14 +569,7 @@ int kobject_uevent_env(struct kobject *kobj, enum kobject_action action,
> }
>
> mutex_lock(&uevent_sock_mutex);
> - /* we will send an event, so request a new sequence number */
> - retval = add_uevent_var(env, "SEQNUM=%llu", (unsigned long long)++uevent_seqnum);
> - if (retval) {
> - mutex_unlock(&uevent_sock_mutex);
> - goto exit;
> - }
> - retval = kobject_uevent_net_broadcast(kobj, env, action_string,
> - devpath);
> + retval = kobject_uevent_net_broadcast(kobj, env, action_string, devpath);
> mutex_unlock(&uevent_sock_mutex);
How does all of this work with events for network devices that are not
in the initial network namespace. This looks to me like this code fails
to take the sk_mutex.
> #ifdef CONFIG_UEVENT_HELPER
> @@ -605,17 +657,18 @@ int add_uevent_var(struct kobj_uevent_env *env, const char *format, ...)
> EXPORT_SYMBOL_GPL(add_uevent_var);
>
> #if defined(CONFIG_NET)
> -static int uevent_net_broadcast(struct sock *usk, struct sk_buff *skb,
> +static int uevent_net_broadcast(struct uevent_sock *ue_sk, struct sk_buff *skb,
> struct netlink_ext_ack *extack)
> {
> - /* u64 to chars: 2^64 - 1 = 21 chars */
> - char buf[sizeof("SEQNUM=") + 21];
> + struct sock *usk = ue_sk->sk;
> + char buf[SEQNUM_BUFSIZE];
> struct sk_buff *skbc;
> int ret;
>
> /* bump and prepare sequence number */
> - ret = snprintf(buf, sizeof(buf), "SEQNUM=%llu", ++uevent_seqnum);
> - if (ret < 0 || (size_t)ret >= sizeof(buf))
> + ret = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu",
> + ++sock_net(ue_sk->sk)->uevent_seqnum);
> + if (ret < 0 || ret >= SEQNUM_BUFSIZE)
> return -ENOMEM;
> ret++;
>
> @@ -668,9 +721,15 @@ static int uevent_net_rcv_skb(struct sk_buff *skb, struct nlmsghdr *nlh,
> return -EPERM;
> }
>
> - mutex_lock(&uevent_sock_mutex);
> - ret = uevent_net_broadcast(net->uevent_sock->sk, skb, extack);
> - mutex_unlock(&uevent_sock_mutex);
> + if (net->user_ns == &init_user_ns)
> + mutex_lock(&uevent_sock_mutex);
> + else
> + mutex_lock(&net->uevent_sock->sk_mutex);
> + ret = uevent_net_broadcast(net->uevent_sock, skb, extack);
> + if (net->user_ns == &init_user_ns)
> + mutex_unlock(&uevent_sock_mutex);
> + else
> + mutex_unlock(&net->uevent_sock->sk_mutex);
>
> return ret;
> }
> @@ -708,6 +767,13 @@ static int uevent_net_init(struct net *net)
> mutex_lock(&uevent_sock_mutex);
> list_add_tail(&ue_sk->list, &uevent_sock_list);
> mutex_unlock(&uevent_sock_mutex);
> + } else {
> + /*
> + * Uevent sockets and counters for network namespaces
> + * not owned by the initial user namespace have their
> + * own mutex.
> + */
> + mutex_init(&ue_sk->sk_mutex);
> }
>
> return 0;
> diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
> index a11e03f920d3..8894638f5150 100644
> --- a/net/core/net_namespace.c
> +++ b/net/core/net_namespace.c
> @@ -618,6 +618,20 @@ struct net *get_net_ns_by_pid(pid_t pid)
> }
> EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
>
> +u64 get_ns_uevent_seqnum_by_vpid(void)
> +{
> + pid_t cur_pid;
> + struct net *net;
> +
> + cur_pid = task_pid_vnr(current);
> + net = get_net_ns_by_pid(cur_pid);
> + if (IS_ERR(net))
> + return 0;
> +
> + return net->uevent_seqnum;
> +}
> +EXPORT_SYMBOL_GPL(get_ns_uevent_seqnum_by_vpid);
I just have to say this function is completely crazy.
You go from the tsk to the pid back to the tsk.
And you leak a struct net pointer.
It is much simpler and less racy to say:
current->nsproxy->net_ns->uevent_seqnum;
That you are accessing current->nsproxy means nsproxy can't
change. The rcu_read_lock etc that get_net_ns_by_pid does
is there for accessing non-current tasks.
> static __net_init int net_ns_net_init(struct net *net)
> {
> #ifdef CONFIG_NET_NS
^ permalink raw reply
* Re: [PATCH bpf-next,v3 0/2] bpf: add helper for getting xfrm states
From: Daniel Borkmann @ 2018-04-24 21:53 UTC (permalink / raw)
To: Eyal Birger, netdev; +Cc: shmulik, ast
In-Reply-To: <1524581430-11921-1-git-send-email-eyal.birger@gmail.com>
On 04/24/2018 04:50 PM, Eyal Birger wrote:
> This patchset adds support for fetching XFRM state information from
> an eBPF program called from TC.
>
> The first patch introduces a helper for fetching an XFRM state from the
> skb's secpath. The XFRM state is modeled using a new virtual struct which
> contains the SPI, peer address, and reqid values of the state; This struct
> can be extended in the future to provide additional state information.
>
> The second patch adds a test example in test_tunnel_bpf.sh. The sample
> validates the correct extraction of state information by the eBPF program.
>
> ---
> v3:
> - Kept SPI and peer IPv4 address in state in network byte order
> following suggestion from Alexei Starovoitov
> v2:
> - Fixed two comments by Daniel Borkmann:
> - disallow reserved flags in helper call
> - avoid compiling in helper code when CONFIG_XFRM is off
Applied to bpf-next, thanks Eyal!
^ permalink raw reply
* Re: [pci PATCH v8 0/4] Add support for unmanaged SR-IOV
From: Bjorn Helgaas @ 2018-04-24 21:51 UTC (permalink / raw)
To: Alexander Duyck
Cc: Alexander Duyck, Bjorn Helgaas, linux-pci, virtio-dev, kvm,
Netdev, Daly, Dan, LKML, linux-nvme, Keith Busch, netanel,
Don Dutile, Maximilian Heyne, Wang, Liang-min, Rustad, Mark D,
David Woodhouse, Christoph Hellwig, dwmw
In-Reply-To: <CAKgT0Ufcq6aQW5zPzBUePO4R_BCa2B_aZ8U2=meOQeOvshZ=Mw@mail.gmail.com>
On Sat, Apr 21, 2018 at 05:22:27PM -0700, Alexander Duyck wrote:
> On Sat, Apr 21, 2018 at 1:34 PM, Bjorn Helgaas <helgaas@kernel.org> wrote:
> > For example, I'm not sure what you mean by "devices where the PF is
> > not capable of managing VF resources."
> >
> > It *sounds* like you're saying the hardware works differently on some
> > devices, but I don't think that's what you mean. I think you're
> > saying something about which drivers are used for the PF and the VF.
>
> That is sort of what I am saying.
>
> So for example with ixgbe there is functionality which is controlled
> in the MMIO space of the PF that affects the functionality of the VFs
> that are generated on the device. The PF has to rearrange the
> resources such as queues and interrupts on the device before it can
> enable SR-IOV, and it could alter those later to limit what the VF is
> capable of doing.
>
> The model I am dealing with via this patch set has a PF that is not
> much different than the VFs other than the fact that it has some
> extended configuration space bits in place for SR-IOV, ARI, ACS, and
> whatever other bits are needed in order to support spawning isolated
> VFs.
OK, thanks for the explanation, I think I understand what's going on
now, correct me if I'm mistaken. I added a hint about "PF" for Randy,
too.
These are on pci/virtualization for v4.18.
commit 8effc395c209
Author: Alexander Duyck <alexander.h.duyck@intel.com>
Date: Sat Apr 21 15:23:09 2018 -0500
PCI/IOV: Add pci_sriov_configure_simple()
SR-IOV (Single Root I/O Virtualization) is an optional PCIe capability (see
PCIe r4.0, sec 9). A PCIe Function with the SR-IOV capability is referred
to as a PF (Physical Function). If SR-IOV is enabled on the PF, several
VFs (Virtual Functions) may be created. The VFs can be individually
assigned to virtual machines, which allows them to share a single hardware
device while being isolated from each other.
Some SR-IOV devices have resources such as queues and interrupts that must
be set up in the PF before enabling the VFs, so they require a PF driver to
do that.
Other SR-IOV devices don't require any PF setup before enabling VFs. Add a
pci_sriov_configure_simple() interface so PF drivers for such devices can
use it without repeating the VF-enabling code.
Tested-by: Mark Rustad <mark.d.rustad@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
[bhelgaas: changelog, comment]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Greg Rose <gvrose8192@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>:wq
commit a8ccf8a66663
Author: Alexander Duyck <alexander.h.duyck@intel.com>
Date: Tue Apr 24 16:47:16 2018 -0500
PCI/IOV: Add pci-pf-stub driver for PFs that only enable VFs
Some SR-IOV PF devices provide no functionality other than acting as a
means of enabling VFs. For these devices, we want to enable the VFs and
assign them to guest virtual machines, but there's no need to have a driver
for the PF itself.
Add a new pci-pf-stub driver to claim those PF devices and provide the
generic VF enable functionality. An administrator can use the sysfs
"sriov_numvfs" file to enable VFs, then assign them to guests.
For now I only have one example ID provided by Amazon in terms of devices
that require this functionality. The general idea is that in the future we
will see other devices added as vendors come up with devices where the PF
is more or less just a lightweight shim used to allocate VFs.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
[bhelgaas: changelog]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Greg Rose <gvrose8192@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
commit 115ddc491922
Author: Alexander Duyck <alexander.h.duyck@intel.com>
Date: Tue Apr 24 16:47:22 2018 -0500
net: ena: Use pci_sriov_configure_simple() to enable VFs
Instead of implementing our own version of a SR-IOV configuration stub in
the ena driver, use the existing pci_sriov_configure_simple() function.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Greg Rose <gvrose8192@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
commit 74d986abc20b
Author: Alexander Duyck <alexander.h.duyck@intel.com>
Date: Tue Apr 24 16:47:27 2018 -0500
nvme-pci: Use pci_sriov_configure_simple() to enable VFs
Instead of implementing our own version of a SR-IOV configuration stub in
the nvme driver, use the existing pci_sriov_configure_simple() function.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
^ permalink raw reply
* [PATCH bpf-next] tools/bpf: remove test_sock_addr from TEST_GEN_PROGS
From: Yonghong Song @ 2018-04-24 21:45 UTC (permalink / raw)
To: ast, daniel, netdev; +Cc: kernel-team
Since test_sock_addr is not supposed to run by itself,
remove it from TEST_GEN_PROGS and add it to
TEST_GEN_PROGS_EXTENDED. This way, run_tests will
not run test_sock_addr. The corresponding test to run
is test_sock_addr.sh.
Signed-off-by: Yonghong Song <yhs@fb.com>
---
tools/testing/selftests/bpf/Makefile | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
index 0b72cc7..64037ee 100644
--- a/tools/testing/selftests/bpf/Makefile
+++ b/tools/testing/selftests/bpf/Makefile
@@ -24,7 +24,7 @@ urandom_read: urandom_read.c
# Order correspond to 'make run_tests' order
TEST_GEN_PROGS = test_verifier test_tag test_maps test_lru_map test_lpm_map test_progs \
test_align test_verifier_log test_dev_cgroup test_tcpbpf_user \
- test_sock test_sock_addr test_btf
+ test_sock test_btf
TEST_GEN_FILES = test_pkt_access.o test_xdp.o test_l4lb.o test_tcp_estats.o test_obj_id.o \
test_pkt_md_access.o test_xdp_redirect.o test_xdp_meta.o sockmap_parse_prog.o \
@@ -43,7 +43,7 @@ TEST_PROGS := test_kmod.sh \
test_sock_addr.sh
# Compile but not part of 'make run_tests'
-TEST_GEN_PROGS_EXTENDED = test_libbpf_open
+TEST_GEN_PROGS_EXTENDED = test_libbpf_open test_sock_addr
include ../lib.mk
--
2.9.5
^ permalink raw reply related
* [PATCH net-next] sctp: fix identification of new acks for SFR-CACC
From: Marcelo Ricardo Leitner @ 2018-04-24 21:17 UTC (permalink / raw)
To: netdev; +Cc: linux-sctp, Xin Long, Vlad Yasevich, Neil Horman
It's currently written as:
if (!tchunk->tsn_gap_acked) { [1]
tchunk->tsn_gap_acked = 1;
...
}
if (TSN_lte(tsn, sack_ctsn)) {
if (!tchunk->tsn_gap_acked) {
/* SFR-CACC processing */
...
}
}
Which causes the SFR-CACC processing on ack reception to never process,
as tchunk->tsn_gap_acked is always true by then. Block [1] was
moved to that position by the commit marked below.
This patch fixes it by doing SFR-CACC processing earlier, before
tsn_gap_acked is set to true.
Fixes: 31b02e154940 ("sctp: Failover transmitted list on transport delete")
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
---
Even though this is a -stable candidate, please apply it to net-next
to avoid conflicts with subsequent patches in my queue. Thanks.
net/sctp/outqueue.c | 48 +++++++++++++++++++++++-------------------------
1 file changed, 23 insertions(+), 25 deletions(-)
diff --git a/net/sctp/outqueue.c b/net/sctp/outqueue.c
index f211b3db6a3543073e113da121bb28518b0af491..dee7cbd5483149024f2f3195db2fe4d473b1a00a 100644
--- a/net/sctp/outqueue.c
+++ b/net/sctp/outqueue.c
@@ -1457,7 +1457,7 @@ static void sctp_check_transmitted(struct sctp_outq *q,
* the outstanding bytes for this chunk, so only
* count bytes associated with a transport.
*/
- if (transport) {
+ if (transport && !tchunk->tsn_gap_acked) {
/* If this chunk is being used for RTT
* measurement, calculate the RTT and update
* the RTO using this value.
@@ -1469,14 +1469,34 @@ static void sctp_check_transmitted(struct sctp_outq *q,
* first instance of the packet or a later
* instance).
*/
- if (!tchunk->tsn_gap_acked &&
- !sctp_chunk_retransmitted(tchunk) &&
+ if (!sctp_chunk_retransmitted(tchunk) &&
tchunk->rtt_in_progress) {
tchunk->rtt_in_progress = 0;
rtt = jiffies - tchunk->sent_at;
sctp_transport_update_rto(transport,
rtt);
}
+
+ if (TSN_lte(tsn, sack_ctsn)) {
+ /*
+ * SFR-CACC algorithm:
+ * 2) If the SACK contains gap acks
+ * and the flag CHANGEOVER_ACTIVE is
+ * set the receiver of the SACK MUST
+ * take the following action:
+ *
+ * B) For each TSN t being acked that
+ * has not been acked in any SACK so
+ * far, set cacc_saw_newack to 1 for
+ * the destination that the TSN was
+ * sent to.
+ */
+ if (sack->num_gap_ack_blocks &&
+ q->asoc->peer.primary_path->cacc.
+ changeover_active)
+ transport->cacc.cacc_saw_newack
+ = 1;
+ }
}
/* If the chunk hasn't been marked as ACKED,
@@ -1508,28 +1528,6 @@ static void sctp_check_transmitted(struct sctp_outq *q,
restart_timer = 1;
forward_progress = true;
- if (!tchunk->tsn_gap_acked) {
- /*
- * SFR-CACC algorithm:
- * 2) If the SACK contains gap acks
- * and the flag CHANGEOVER_ACTIVE is
- * set the receiver of the SACK MUST
- * take the following action:
- *
- * B) For each TSN t being acked that
- * has not been acked in any SACK so
- * far, set cacc_saw_newack to 1 for
- * the destination that the TSN was
- * sent to.
- */
- if (transport &&
- sack->num_gap_ack_blocks &&
- q->asoc->peer.primary_path->cacc.
- changeover_active)
- transport->cacc.cacc_saw_newack
- = 1;
- }
-
list_add_tail(&tchunk->transmitted_list,
&q->sacked);
} else {
^ permalink raw reply related
* [PATCH net-next] sctp: fix const parameter violation in sctp_make_sack
From: Marcelo Ricardo Leitner @ 2018-04-24 21:17 UTC (permalink / raw)
To: netdev; +Cc: linux-sctp, Xin Long, Vlad Yasevich, Neil Horman
sctp_make_sack() make changes to the asoc and this cast is just
bypassing the const attribute. As there is no need to have the const
there, just remove it and fix the violation.
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
---
This one can go to net or net-next, but targetting net-next here just to
keep it together with the rest (which I'll post as patches get in).
include/net/sctp/sm.h | 2 +-
net/sctp/sm_make_chunk.c | 9 ++++-----
2 files changed, 5 insertions(+), 6 deletions(-)
diff --git a/include/net/sctp/sm.h b/include/net/sctp/sm.h
index 2d0e782c90551377ad654bcef1224bbdb75ba394..f4b657478a304050851f33d92c71162a4a4a2e50 100644
--- a/include/net/sctp/sm.h
+++ b/include/net/sctp/sm.h
@@ -207,7 +207,7 @@ struct sctp_chunk *sctp_make_datafrag_empty(const struct sctp_association *asoc,
int len, __u8 flags, gfp_t gfp);
struct sctp_chunk *sctp_make_ecne(const struct sctp_association *asoc,
const __u32 lowest_tsn);
-struct sctp_chunk *sctp_make_sack(const struct sctp_association *asoc);
+struct sctp_chunk *sctp_make_sack(struct sctp_association *asoc);
struct sctp_chunk *sctp_make_shutdown(const struct sctp_association *asoc,
const struct sctp_chunk *chunk);
struct sctp_chunk *sctp_make_shutdown_ack(const struct sctp_association *asoc,
diff --git a/net/sctp/sm_make_chunk.c b/net/sctp/sm_make_chunk.c
index 5a4fb1dc8400a0316177ce65be8126857297eb5e..db93eabd6ef500ab20be6e091ee06bd3b60923c9 100644
--- a/net/sctp/sm_make_chunk.c
+++ b/net/sctp/sm_make_chunk.c
@@ -779,10 +779,9 @@ struct sctp_chunk *sctp_make_datafrag_empty(const struct sctp_association *asoc,
* association. This reports on which TSN's we've seen to date,
* including duplicates and gaps.
*/
-struct sctp_chunk *sctp_make_sack(const struct sctp_association *asoc)
+struct sctp_chunk *sctp_make_sack(struct sctp_association *asoc)
{
struct sctp_tsnmap *map = (struct sctp_tsnmap *)&asoc->peer.tsn_map;
- struct sctp_association *aptr = (struct sctp_association *)asoc;
struct sctp_gap_ack_block gabs[SCTP_MAX_GABS];
__u16 num_gabs, num_dup_tsns;
struct sctp_transport *trans;
@@ -857,7 +856,7 @@ struct sctp_chunk *sctp_make_sack(const struct sctp_association *asoc)
/* Add the duplicate TSN information. */
if (num_dup_tsns) {
- aptr->stats.idupchunks += num_dup_tsns;
+ asoc->stats.idupchunks += num_dup_tsns;
sctp_addto_chunk(retval, sizeof(__u32) * num_dup_tsns,
sctp_tsnmap_get_dups(map));
}
@@ -869,11 +868,11 @@ struct sctp_chunk *sctp_make_sack(const struct sctp_association *asoc)
* association so no transport will match after a wrap event like this,
* Until the next sack
*/
- if (++aptr->peer.sack_generation == 0) {
+ if (++asoc->peer.sack_generation == 0) {
list_for_each_entry(trans, &asoc->peer.transport_addr_list,
transports)
trans->sack_generation = 0;
- aptr->peer.sack_generation = 1;
+ asoc->peer.sack_generation = 1;
}
nodata:
return retval;
^ permalink raw reply related
* [PATCH net-next] neighbour: support for NTF_EXT_LEARNED flag
From: Roopa Prabhu @ 2018-04-24 20:49 UTC (permalink / raw)
To: davem; +Cc: netdev, nikolay, dsa
From: Roopa Prabhu <roopa@cumulusnetworks.com>
This patch extends NTF_EXT_LEARNED support to the neighbour system.
Example use-case: An Ethernet VPN implementation (eg in FRR routing suite)
can use this flag to add dynamic reachable external neigh entires
learned via control plane. The use of neigh NTF_EXT_LEARNED in this
patch is consistent with its use with bridge and vxlan fdb entries.
Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
---
include/net/neighbour.h | 19 ++++++++++++++++++-
net/core/neighbour.c | 8 +++++++-
2 files changed, 25 insertions(+), 2 deletions(-)
diff --git a/include/net/neighbour.h b/include/net/neighbour.h
index e421f86..6c1eecd 100644
--- a/include/net/neighbour.h
+++ b/include/net/neighbour.h
@@ -246,6 +246,7 @@ static inline void *neighbour_priv(const struct neighbour *n)
#define NEIGH_UPDATE_F_OVERRIDE 0x00000001
#define NEIGH_UPDATE_F_WEAK_OVERRIDE 0x00000002
#define NEIGH_UPDATE_F_OVERRIDE_ISROUTER 0x00000004
+#define NEIGH_UPDATE_F_EXT_LEARNED 0x20000000
#define NEIGH_UPDATE_F_ISROUTER 0x40000000
#define NEIGH_UPDATE_F_ADMIN 0x80000000
@@ -526,5 +527,21 @@ static inline void neigh_ha_snapshot(char *dst, const struct neighbour *n,
} while (read_seqretry(&n->ha_lock, seq));
}
-
+static inline void neigh_update_ext_learned(struct neighbour *neigh, u32 flags,
+ int *notify)
+{
+ u8 ndm_flags = 0;
+
+ if (!(flags & NEIGH_UPDATE_F_ADMIN))
+ return;
+
+ ndm_flags |= (flags & NEIGH_UPDATE_F_EXT_LEARNED) ? NTF_EXT_LEARNED : 0;
+ if ((neigh->flags ^ ndm_flags) & NTF_EXT_LEARNED) {
+ if (ndm_flags & NTF_EXT_LEARNED)
+ neigh->flags |= NTF_EXT_LEARNED;
+ else
+ neigh->flags &= ~NTF_EXT_LEARNED;
+ *notify = 1;
+ }
+}
#endif
diff --git a/net/core/neighbour.c b/net/core/neighbour.c
index ce51986..5afae29 100644
--- a/net/core/neighbour.c
+++ b/net/core/neighbour.c
@@ -820,7 +820,8 @@ static void neigh_periodic_work(struct work_struct *work)
write_lock(&n->lock);
state = n->nud_state;
- if (state & (NUD_PERMANENT | NUD_IN_TIMER)) {
+ if ((state & (NUD_PERMANENT | NUD_IN_TIMER)) ||
+ (n->flags & NTF_EXT_LEARNED)) {
write_unlock(&n->lock);
goto next_elt;
}
@@ -1136,6 +1137,8 @@ int neigh_update(struct neighbour *neigh, const u8 *lladdr, u8 new,
if (neigh->dead)
goto out;
+ neigh_update_ext_learned(neigh, flags, ¬ify);
+
if (!(new & NUD_VALID)) {
neigh_del_timer(neigh);
if (old & NUD_CONNECTED)
@@ -1781,6 +1784,9 @@ static int neigh_add(struct sk_buff *skb, struct nlmsghdr *nlh,
flags &= ~NEIGH_UPDATE_F_OVERRIDE;
}
+ if (ndm->ndm_flags & NTF_EXT_LEARNED)
+ flags |= NEIGH_UPDATE_F_EXT_LEARNED;
+
if (ndm->ndm_flags & NTF_USE) {
neigh_event_send(neigh, NULL);
err = 0;
--
2.1.4
^ permalink raw reply related
* Re: [PATCH net-next 2/2 v1] netns: isolate seqnums to use per-netns locks
From: Christian Brauner @ 2018-04-24 20:45 UTC (permalink / raw)
To: David Miller
Cc: ebiederm, netdev, linux-kernel, avagin, ktkhai, serge, gregkh
In-Reply-To: <20180424.153925.2063217911734444324.davem@davemloft.net>
On Tue, Apr 24, 2018 at 03:39:25PM -0400, David Miller wrote:
> From: Christian Brauner <christian.brauner@ubuntu.com>
> Date: Mon, 23 Apr 2018 12:24:43 +0200
>
> > + #ifdef CONFIG_NET
> > + seqnum = get_ns_uevent_seqnum_by_vpid();
> > + #else
> > + seqnum = uevent_seqnum;
> > + #endif
>
> Please don't indend the code like this.
>
> By indenting the CPP directives, which should be at column zero, the
> actual code became double indented.
Ah, sorry. Sent v2 with the indendation fixed just now.
Thanks!
Christian
^ permalink raw reply
* [PATCH net-next 2/2 v2] netns: isolate seqnums to use per-netns locks
From: Christian Brauner @ 2018-04-24 20:43 UTC (permalink / raw)
To: ebiederm, davem, netdev, linux-kernel
Cc: avagin, ktkhai, serge, gregkh, Christian Brauner
In-Reply-To: <20180424204335.12904-1-christian.brauner@ubuntu.com>
Now that it's possible to have a different set of uevents in different
network namespaces, per-network namespace uevent sequence numbers are
introduced. This increases performance as locking is now restricted to the
network namespace affected by the uevent rather than locking everything.
Testing revealed significant performance improvements. For details see
"Testing" below.
Since commit 692ec06 ("netns: send uevent messages") network namespaces not
owned by the intial user namespace can be sent uevents from a sufficiently
privileged userspace process.
In order to send a uevent into a network namespace not owned by the initial
user namespace we currently still need to take the *global mutex* that
locks the uevent socket list even though the list *only contains network
namespaces owned by the initial user namespace*. This needs to be done
because the uevent counter is a global variable. Taking the global lock is
performance sensitive since a user on the host can spawn a pool of n
process that each create their own new user and network namespaces and then
go on to inject uevents in parallel into the network namespace of all of
these processes. This can have a significant performance impact for the
host's udevd since it means that there can be a lot of delay between a
device being added and the corresponding uevent being sent out and
available for processing by udevd. It also means that each network
namespace not owned by the initial user namespace which userspace has sent
a uevent to will need to wait until the lock becomes available.
Implementation:
This patch gives each network namespace its own uevent sequence number.
Each network namespace not owned by the initial user namespace receives its
own mutex. The struct uevent_sock is opaque to callers outside of kobject.c
so the mutex *can* and *is* only ever accessed in lib/kobject.c. In this
file it is clearly documented which lock has to be taken. All network
namespaces owned by the initial user namespace will still share the same
lock since they are all served sequentially via the uevent socket list.
This decouples the locking and ensures that the host retrieves uevents as
fast as possible even if there are a lot of uevents injected into network
namespaces not owned by the initial user namespace. In addition, each
network namespace not owned by the initial user namespace does not have to
wait on any other network namespace not sharing the same user namespace.
Testing:
Two 4.17-rc1 test kernels were compiled. One with per netns uevent seqnums
with decoupled locking and one without. To ensure that testing made sense
both kernels carried the patch to remove network namespaces not owned by
the initial user namespace from the uevent socket list.
Three tests were constructed. All of them showed significant performance
improvements with per-netns uevent sequence numbers and decoupled locking.
# Testcase 1:
Only Injecting Uevents into network namespaces not owned by the initial
user namespace.
- created 1000 new user namespace + network namespace pairs
- opened a uevent listener in each of those namespace pairs
- injected uevents into each of those network namespaces 10,000 times
meaning 10,000,000 (10 million) uevents were injected. (The high
number of uevent injections should get rid of a lot of jitter.)
The injection was done by fork()ing 1000 uevent injectors in a simple
for-loop to ensure that uevents were injected in parallel.
- mean transaction time was calculated:
- *without* uevent sequence number namespacing: 67 μs
- *with* uevent sequence number namespacing: 55 μs
- makes a difference of: 12 μs
- a t-test was performed on the two data vectors which revealed
shows significant performance improvements:
Welch Two Sample t-test
data: x1 and y1
t = 405.16, df = 18883000, p-value < 2.2e-16
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
12.14949 12.26761
sample estimates:
mean of x mean of y
68.48594 56.27739
# Testcase 2:
Injecting Uevents into network namespaces not owned by the initial user
namespace and network namespaces owned by the initial user namespace.
- created 500 new user namespace + network namespace pairs
- created 500 new network namespace pairs
- opened a uevent listener in each of those namespace pairs
- injected uevents into each of those network namespaces 10,000 times
meaning 10,000,000 (10 million) uevents were injected. (The high
number of uevent injections should get rid of a lot of jitter.)
The injection was done by fork()ing 1000 uevent injectors in a simple
for-loop to ensure that uevents were injected in parallel.
- mean transaction time was calculated:
- *without* uevent sequence number namespacing: 572 μs
- *with* uevent sequence number namespacing: 514 μs
- makes a difference of: 58 μs
- a t-test was performed on the two data vectors which revealed
shows significant performance improvements:
Welch Two Sample t-test
data: x2 and y2
t = 38.685, df = 19682000, p-value < 2.2e-16
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
55.10630 60.98815
sample estimates:
mean of x mean of y
572.9684 514.9211
# Testcase 3:
Created 500 new user namespace + network namespace pairs *without uevent
listeners*
- created 500 new network namespace pairs *without uevent listeners*
- injected uevents into each of those network namespaces 10,000 times
meaning 10,000,000 (10 million) uevents were injected. (The high number
of uevent injections should get rid of a lot of jitter.)
The injection was done by fork()ing 1000 uevent injectors in a simple
for-loop to ensure that uevents were injected in parallel.
- mean transaction time was calculated:
- *without* uevent sequence number namespacing: 206 μs
- *with* uevent sequence number namespacing: 163 μs
- makes a difference of: 43 μs
- a t-test was performed on the two data vectors which revealed
shows significant performance improvements:
Welch Two Sample t-test
data: x3 and y3
t = 58.37, df = 17711000, p-value < 2.2e-16
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
41.77860 44.68178
sample estimates:
mean of x mean of y
207.2632 164.0330
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
---
Changelog v1->v2:
* non-functional change: fix indendation for C directives in
kernel/ksysfs.c
Changelog v0->v1:
* add detailed test results to the commit message
* account for kernels compiled without CONFIG_NET
---
include/linux/kobject.h | 2 +
include/net/net_namespace.h | 3 ++
kernel/ksysfs.c | 11 +++-
lib/kobject_uevent.c | 104 +++++++++++++++++++++++++++++-------
net/core/net_namespace.c | 14 +++++
5 files changed, 114 insertions(+), 20 deletions(-)
diff --git a/include/linux/kobject.h b/include/linux/kobject.h
index 7f6f93c3df9c..4e608968907f 100644
--- a/include/linux/kobject.h
+++ b/include/linux/kobject.h
@@ -36,8 +36,10 @@
extern char uevent_helper[];
#endif
+#ifndef CONFIG_NET
/* counter to tag the uevent, read only except for the kobject core */
extern u64 uevent_seqnum;
+#endif
/*
* The actions here must match the index to the string array
diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
index 47e35cce3b64..e4e171b1ba69 100644
--- a/include/net/net_namespace.h
+++ b/include/net/net_namespace.h
@@ -85,6 +85,8 @@ struct net {
struct sock *genl_sock;
struct uevent_sock *uevent_sock; /* uevent socket */
+ /* counter to tag the uevent, read only except for the kobject core */
+ u64 uevent_seqnum;
struct list_head dev_base_head;
struct hlist_head *dev_name_head;
@@ -189,6 +191,7 @@ extern struct list_head net_namespace_list;
struct net *get_net_ns_by_pid(pid_t pid);
struct net *get_net_ns_by_fd(int fd);
+u64 get_ns_uevent_seqnum_by_vpid(void);
#ifdef CONFIG_SYSCTL
void ipx_register_sysctl(void);
diff --git a/kernel/ksysfs.c b/kernel/ksysfs.c
index 46ba853656f6..38b70b90a21f 100644
--- a/kernel/ksysfs.c
+++ b/kernel/ksysfs.c
@@ -19,6 +19,7 @@
#include <linux/sched.h>
#include <linux/capability.h>
#include <linux/compiler.h>
+#include <net/net_namespace.h>
#include <linux/rcupdate.h> /* rcu_expedited and rcu_normal */
@@ -33,7 +34,15 @@ static struct kobj_attribute _name##_attr = \
static ssize_t uevent_seqnum_show(struct kobject *kobj,
struct kobj_attribute *attr, char *buf)
{
- return sprintf(buf, "%llu\n", (unsigned long long)uevent_seqnum);
+ u64 seqnum;
+
+#ifdef CONFIG_NET
+ seqnum = get_ns_uevent_seqnum_by_vpid();
+#else
+ seqnum = uevent_seqnum;
+#endif
+
+ return sprintf(buf, "%llu\n", (unsigned long long)seqnum);
}
KERNEL_ATTR_RO(uevent_seqnum);
diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
index f5f5038787ac..5da20def556d 100644
--- a/lib/kobject_uevent.c
+++ b/lib/kobject_uevent.c
@@ -29,21 +29,42 @@
#include <net/net_namespace.h>
+#ifndef CONFIG_NET
u64 uevent_seqnum;
+#endif
+
#ifdef CONFIG_UEVENT_HELPER
char uevent_helper[UEVENT_HELPER_PATH_LEN] = CONFIG_UEVENT_HELPER_PATH;
#endif
+/*
+ * Size a buffer needs to be in order to hold the largest possible sequence
+ * number stored in a u64 including \0 byte: 2^64 - 1 = 21 chars.
+ */
+#define SEQNUM_BUFSIZE (sizeof("SEQNUM=") + 21)
struct uevent_sock {
struct list_head list;
struct sock *sk;
+ /*
+ * This mutex protects uevent sockets and the uevent counter of
+ * network namespaces *not* owned by init_user_ns.
+ * For network namespaces owned by init_user_ns this lock is *not*
+ * valid instead the global uevent_sock_mutex must be used!
+ */
+ struct mutex sk_mutex;
};
#ifdef CONFIG_NET
static LIST_HEAD(uevent_sock_list);
#endif
-/* This lock protects uevent_seqnum and uevent_sock_list */
+/*
+ * This mutex protects uevent sockets and the uevent counter of network
+ * namespaces owned by init_user_ns.
+ * For network namespaces not owned by init_user_ns this lock is *not*
+ * valid instead the network namespace specific sk_mutex in struct
+ * uevent_sock must be used!
+ */
static DEFINE_MUTEX(uevent_sock_mutex);
/* the strings here must match the enum in include/linux/kobject.h */
@@ -253,6 +274,22 @@ static int kobj_bcast_filter(struct sock *dsk, struct sk_buff *skb, void *data)
return 0;
}
+
+static bool can_hold_seqnum(const struct kobj_uevent_env *env, size_t len)
+{
+ if (env->envp_idx >= ARRAY_SIZE(env->envp)) {
+ WARN(1, KERN_ERR "Failed to append sequence number. "
+ "Too many uevent variables\n");
+ return false;
+ }
+
+ if ((env->buflen + len) > UEVENT_BUFFER_SIZE) {
+ WARN(1, KERN_ERR "Insufficient space to append sequence number\n");
+ return false;
+ }
+
+ return true;
+}
#endif
#ifdef CONFIG_UEVENT_HELPER
@@ -308,18 +345,22 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
/* send netlink message */
list_for_each_entry(ue_sk, &uevent_sock_list, list) {
+ /* bump sequence number */
+ u64 seqnum = ++sock_net(ue_sk->sk)->uevent_seqnum;
struct sock *uevent_sock = ue_sk->sk;
+ char buf[SEQNUM_BUFSIZE];
if (!netlink_has_listeners(uevent_sock, 1))
continue;
if (!skb) {
- /* allocate message with the maximum possible size */
+ /* calculate header length */
size_t len = strlen(action_string) + strlen(devpath) + 2;
char *scratch;
+ /* allocate message with the maximum possible size */
retval = -ENOMEM;
- skb = alloc_skb(len + env->buflen, GFP_KERNEL);
+ skb = alloc_skb(len + env->buflen + SEQNUM_BUFSIZE, GFP_KERNEL);
if (!skb)
continue;
@@ -327,11 +368,24 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
scratch = skb_put(skb, len);
sprintf(scratch, "%s@%s", action_string, devpath);
+ /* add env */
skb_put_data(skb, env->buf, env->buflen);
NETLINK_CB(skb).dst_group = 1;
}
+ /* prepare netns seqnum */
+ retval = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu", seqnum);
+ if (retval < 0 || retval >= SEQNUM_BUFSIZE)
+ continue;
+ retval++;
+
+ if (!can_hold_seqnum(env, retval))
+ continue;
+
+ /* append netns seqnum */
+ skb_put_data(skb, buf, retval);
+
retval = netlink_broadcast_filtered(uevent_sock, skb_get(skb),
0, 1, GFP_KERNEL,
kobj_bcast_filter,
@@ -339,8 +393,13 @@ static int kobject_uevent_net_broadcast(struct kobject *kobj,
/* ENOBUFS should be handled in userspace */
if (retval == -ENOBUFS || retval == -ESRCH)
retval = 0;
+
+ /* remove netns seqnum */
+ skb_trim(skb, env->buflen);
}
consume_skb(skb);
+#else
+ uevent_seqnum++;
#endif
return retval;
}
@@ -510,14 +569,7 @@ int kobject_uevent_env(struct kobject *kobj, enum kobject_action action,
}
mutex_lock(&uevent_sock_mutex);
- /* we will send an event, so request a new sequence number */
- retval = add_uevent_var(env, "SEQNUM=%llu", (unsigned long long)++uevent_seqnum);
- if (retval) {
- mutex_unlock(&uevent_sock_mutex);
- goto exit;
- }
- retval = kobject_uevent_net_broadcast(kobj, env, action_string,
- devpath);
+ retval = kobject_uevent_net_broadcast(kobj, env, action_string, devpath);
mutex_unlock(&uevent_sock_mutex);
#ifdef CONFIG_UEVENT_HELPER
@@ -605,17 +657,18 @@ int add_uevent_var(struct kobj_uevent_env *env, const char *format, ...)
EXPORT_SYMBOL_GPL(add_uevent_var);
#if defined(CONFIG_NET)
-static int uevent_net_broadcast(struct sock *usk, struct sk_buff *skb,
+static int uevent_net_broadcast(struct uevent_sock *ue_sk, struct sk_buff *skb,
struct netlink_ext_ack *extack)
{
- /* u64 to chars: 2^64 - 1 = 21 chars */
- char buf[sizeof("SEQNUM=") + 21];
+ struct sock *usk = ue_sk->sk;
+ char buf[SEQNUM_BUFSIZE];
struct sk_buff *skbc;
int ret;
/* bump and prepare sequence number */
- ret = snprintf(buf, sizeof(buf), "SEQNUM=%llu", ++uevent_seqnum);
- if (ret < 0 || (size_t)ret >= sizeof(buf))
+ ret = snprintf(buf, SEQNUM_BUFSIZE, "SEQNUM=%llu",
+ ++sock_net(ue_sk->sk)->uevent_seqnum);
+ if (ret < 0 || ret >= SEQNUM_BUFSIZE)
return -ENOMEM;
ret++;
@@ -668,9 +721,15 @@ static int uevent_net_rcv_skb(struct sk_buff *skb, struct nlmsghdr *nlh,
return -EPERM;
}
- mutex_lock(&uevent_sock_mutex);
- ret = uevent_net_broadcast(net->uevent_sock->sk, skb, extack);
- mutex_unlock(&uevent_sock_mutex);
+ if (net->user_ns == &init_user_ns)
+ mutex_lock(&uevent_sock_mutex);
+ else
+ mutex_lock(&net->uevent_sock->sk_mutex);
+ ret = uevent_net_broadcast(net->uevent_sock, skb, extack);
+ if (net->user_ns == &init_user_ns)
+ mutex_unlock(&uevent_sock_mutex);
+ else
+ mutex_unlock(&net->uevent_sock->sk_mutex);
return ret;
}
@@ -708,6 +767,13 @@ static int uevent_net_init(struct net *net)
mutex_lock(&uevent_sock_mutex);
list_add_tail(&ue_sk->list, &uevent_sock_list);
mutex_unlock(&uevent_sock_mutex);
+ } else {
+ /*
+ * Uevent sockets and counters for network namespaces
+ * not owned by the initial user namespace have their
+ * own mutex.
+ */
+ mutex_init(&ue_sk->sk_mutex);
}
return 0;
diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
index a11e03f920d3..8894638f5150 100644
--- a/net/core/net_namespace.c
+++ b/net/core/net_namespace.c
@@ -618,6 +618,20 @@ struct net *get_net_ns_by_pid(pid_t pid)
}
EXPORT_SYMBOL_GPL(get_net_ns_by_pid);
+u64 get_ns_uevent_seqnum_by_vpid(void)
+{
+ pid_t cur_pid;
+ struct net *net;
+
+ cur_pid = task_pid_vnr(current);
+ net = get_net_ns_by_pid(cur_pid);
+ if (IS_ERR(net))
+ return 0;
+
+ return net->uevent_seqnum;
+}
+EXPORT_SYMBOL_GPL(get_ns_uevent_seqnum_by_vpid);
+
static __net_init int net_ns_net_init(struct net *net)
{
#ifdef CONFIG_NET_NS
--
2.17.0
^ permalink raw reply related
* [PATCH net-next 1/2 v2] netns: restrict uevents
From: Christian Brauner @ 2018-04-24 20:43 UTC (permalink / raw)
To: ebiederm, davem, netdev, linux-kernel
Cc: avagin, ktkhai, serge, gregkh, Christian Brauner
In-Reply-To: <20180424204335.12904-1-christian.brauner@ubuntu.com>
commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
enabled sending hotplug events into all network namespaces back in 2010.
Over time the set of uevents that get sent into all network namespaces has
shrunk a little. We have now reached the point where hotplug events for all
devices that carry a namespace tag are filtered according to that
namespace. Specifically, they are filtered whenever the namespace tag of
the kobject does not match the namespace tag of the netlink socket. One
example are network devices. Uevents for network devices only show up in
the network namespaces these devices are moved to or created in.
However, any uevent for a kobject that does not have a namespace tag
associated with it will not be filtered and we will broadcast it into all
network namespaces. This behavior stopped making sense when user namespaces
were introduced.
This patch restricts uevents to the initial user namespace for a couple of
reasons that have been extensively discusses on the mailing list [1].
- Thundering herd:
Broadcasting uevents into all network namespaces introduces significant
overhead.
All processes that listen to uevents running in non-initial user
namespaces will end up responding to uevents that will be meaningless to
them. Mainly, because non-initial user namespaces cannot easily manage
devices unless they have a privileged host-process helping them out. This
means that there will be a thundering herd of activity when there
shouldn't be any.
- Uevents from non-root users are already filtered in userspace:
Uevents are filtered by userspace in a user namespace because the
received uid != 0. Instead the uid associated with the event will be
65534 == "nobody" because the global root uid is not mapped.
This means we can safely and without introducing regressions modify the
kernel to not send uevents into all network namespaces whose owning user
namespace is not the initial user namespace because we know that
userspace will ignore the message because of the uid anyway. I have
a) verified that is is true for every udev implementation out there b)
that this behavior has been present in all udev implementations from the
very beginning.
- Removing needless overhead/Increasing performance:
Currently, the uevent socket for each network namespace is added to the
global variable uevent_sock_list. The list itself needs to be protected
by a mutex. So everytime a uevent is generated the mutex is taken on the
list. The mutex is held *from the creation of the uevent (memory
allocation, string creation etc. until all uevent sockets have been
handled*. This is aggravated by the fact that for each uevent socket that
has listeners the mc_list must be walked as well which means we're
talking O(n^2) here. Given that a standard Linux workload usually has
quite a lot of network namespaces and - in the face of containers - a lot
of user namespaces this quickly becomes a performance problem (see
"Thundering herd" above). By just recording uevent sockets of network
namespaces that are owned by the initial user namespace we significantly
increase performance in this codepath.
- Injecting uevents:
There's a valid argument that containers might be interested in receiving
device events especially if they are delegated to them by a privileged
userspace process. One prime example are SR-IOV enabled devices that are
explicitly designed to be handed of to other users such as VMs or
containers.
This use-case can now be correctly handled since
commit 692ec06d7c92 ("netns: send uevent messages"). This commit
introduced the ability to send uevents from userspace. As such we can let
a sufficiently privileged (CAP_SYS_ADMIN in the owning user namespace of
the network namespace of the netlink socket) userspace process make a
decision what uevents should be sent. This removes the need to blindly
broadcast uevents into all user namespaces and provides a performant and
safe solution to this problem.
- Filtering logic:
This patch filters by *owning user namespace of the network namespace a
given task resides in* and not by user namespace of the task per se. This
means if the user namespace of a given task is unshared but the network
namespace is kept and is owned by the initial user namespace a listener
that is opening the uevent socket in that network namespace can still
listen to uevents.
[1]: https://lkml.org/lkml/2018/4/4/739
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
---
Changelog v1->v2:
* patch unchanged
Changelog v0->v1:
* patch unchanged
---
lib/kobject_uevent.c | 18 ++++++++++++------
1 file changed, 12 insertions(+), 6 deletions(-)
diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
index 15ea216a67ce..f5f5038787ac 100644
--- a/lib/kobject_uevent.c
+++ b/lib/kobject_uevent.c
@@ -703,9 +703,13 @@ static int uevent_net_init(struct net *net)
net->uevent_sock = ue_sk;
- mutex_lock(&uevent_sock_mutex);
- list_add_tail(&ue_sk->list, &uevent_sock_list);
- mutex_unlock(&uevent_sock_mutex);
+ /* Restrict uevents to initial user namespace. */
+ if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
+ mutex_lock(&uevent_sock_mutex);
+ list_add_tail(&ue_sk->list, &uevent_sock_list);
+ mutex_unlock(&uevent_sock_mutex);
+ }
+
return 0;
}
@@ -713,9 +717,11 @@ static void uevent_net_exit(struct net *net)
{
struct uevent_sock *ue_sk = net->uevent_sock;
- mutex_lock(&uevent_sock_mutex);
- list_del(&ue_sk->list);
- mutex_unlock(&uevent_sock_mutex);
+ if (sock_net(ue_sk->sk)->user_ns == &init_user_ns) {
+ mutex_lock(&uevent_sock_mutex);
+ list_del(&ue_sk->list);
+ mutex_unlock(&uevent_sock_mutex);
+ }
netlink_kernel_release(ue_sk->sk);
kfree(ue_sk);
--
2.17.0
^ permalink raw reply related
page: next (older) | prev (newer) | latest
- recent:[subjects (threaded)|topics (new)|topics (active)]
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox