netfilter: IDLETIMER target v1 - match Android layout
Subsystems (3)
PATCH 1/1 netfilter: IDLETIMER target v1 - match Android layout
From: Maciej Żenczykowski <[email protected]>
Android has long had an extension to IDLETIMER to send netlink messages to userspace, see:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/include/uapi/linux/netfilter/xt_IDLETIMER.h#42
Note: this is idletimer target rev 1, there is no rev 0 in the Android common kernel sources, see registration at:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/net/netfilter/xt_IDLETIMER.c#483
When we compare that to upstream's new idletimer target rev 1:
https://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf-next.git/tree/include/uapi/linux/netfilter/xt_IDLETIMER.h#n46
We immediately notice that these two rev 1 structs are the same size and layout, and that while timer_type and send_nl_msg are differently named and serve a different purpose, they're at the same offset.
This makes them impossible to tell apart - and thus one cannot know in a mixed Android/vanilla environment whether one means timer_type or send_nl_msg.
Since this is iptables/netfilter uapi it introduces a problem between iptables (vanilla vs Android) userspace and kernel (vanilla vs Android) if the two don't match each other.
Additionally when at some point in the future Android picks up 5.7+ it's not at all clear how to resolve the resulting merge conflict.
Furthermore, since upgrading the kernel on old Android phones is pretty much impossible there does not seem to be an easy way out of this predicament.
The only thing I've been able to come up with is some super disgusting kernel version >= 5.7 check in the iptables binary to flip between different struct layouts.
By adding a dummy field to the vanilla Linux kernel header file we can force the two structs to be compatible with each other.
Long term I think I would like to deprecate send_nl_msg out of Android entirely, but I haven't quite been able to figure out exactly how we depend on it. It seems to be very similar to sysfs notifications but with some extra info.
Currently it's actually always enabled whenever Android uses the IDLETIMER target, so we could also probably entirely remove it from the uapi in favour of just always enabling it, but again we can't upgrade old kernels already in the field.
(Also note that this doesn't change the structure's size, as it is simply fitting into the pre-existing padding, and that since 5.7 hasn't been released yet, there's still time to make this uapi visible change)
--- include/uapi/linux/netfilter/xt_IDLETIMER.h | 1 + 1 file changed, 1 insertion(+)
include/uapi/linux/netfilter/xt_IDLETIMER.h +1 -0
@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */ __u8 timer_type;/* for kernel module internal use only */
Signature · 1 line
2.26.0.rc2.310.g2932bb562d-goog
Pablo Neira Ayuso
maintainerQuoted 68 lines · Maciej Żenczykowski
From: Maciej Żenczykowski <[email protected]>
Android has long had an extension to IDLETIMER to send netlink messages to userspace, see:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/include/uapi/linux/netfilter/xt_IDLETIMER.h#42
Note: this is idletimer target rev 1, there is no rev 0 in the Android common kernel sources, see registration at:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/net/netfilter/xt_IDLETIMER.c#483When we compare that to upstream's new idletimer target rev 1:
https://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf-next.git/tree/include/uapi/linux/netfilter/xt_IDLETIMER.h#n46We immediately notice that these two rev 1 structs are the same size and layout, and that while
timer_typeandsend_nl_msgare differently named and serve a different purpose, they're at the same offset.This makes them impossible to tell apart - and thus one cannot know in a mixed Android/vanilla environment whether one means
timer_typeorsend_nl_msg.Since this is iptables/netfilter uapi it introduces a problem between iptables (vanilla vs Android) userspace and kernel (vanilla vs Android) if the two don't match each other.
Additionally when at some point in the future Android picks up 5.7+ it's not at all clear how to resolve the resulting merge conflict.
Furthermore, since upgrading the kernel on old Android phones is pretty much impossible there does not seem to be an easy way out of this predicament.
The only thing I've been able to come up with is some super disgusting kernel version >= 5.7 check in the iptables binary to flip between different struct layouts.
By adding a dummy field to the vanilla Linux kernel header file we can force the two structs to be compatible with each other.
Long term I think I would like to deprecate
send_nl_msgout of Android entirely, but I haven't quite been able to figure out exactly how we depend on it. It seems to be very similar to sysfs notifications but with some extra info.Currently it's actually always enabled whenever Android uses the IDLETIMER target, so we could also probably entirely remove it from the uapi in favour of just always enabling it, but again we can't upgrade old kernels already in the field.
(Also note that this doesn't change the structure's size, as it is simply fitting into the pre-existing padding, and that since 5.7 hasn't been released yet, there's still time to make this uapi visible change)
Cc Manoj Basapathi <[email protected]>Cc Subash Abhinov Kasiviswanathan <[email protected]>Cc Pablo Neira Ayuso <[email protected]>Signed-off-by Maciej Żenczykowski <[email protected]>--- include/uapi/linux/netfilter/xt_IDLETIMER.h | 1 + 1 file changed, 1 insertion(+)include/uapi/linux/netfilter/xt_IDLETIMER.h +0
-0
include/uapi/linux/netfilter/xt_IDLETIMER.h struct idletimer_tg_info_v1 { · line 51
Quoted 4 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h +1
-0@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */
Please, add client code for this send_nl_msg field.
Thank you.
By client code do you mean code for the iptables userspace binary?
Pablo Neira Ayuso
maintainerQuoted 1 line · Maciej Żenczykowski
By client code do you mean code for the iptables userspace binary?
Code that uses this, please.
Quoted 11 lines · Maciej Żenczykowski
Signed-off-by Maciej Żenczykowski <[email protected]>---
include/uapi/linux/netfilter/xt_IDLETIMER.h| 1 +1 file changed, 1 insertion(+)include/uapi/linux/netfilter/xt_IDLETIMER.h +0
-0
include/uapi/linux/netfilter/xt_IDLETIMER.h struct idletimer_tg_info_v1 { · line 54
Quoted 7 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h +1
-0@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */ __u8 timer_type;/* for kernel module internal use only */
This breaks the ABI for law-abiding Linux users (i.e. the GNU/Linux subgroup of it), which is equally terrible.
You will have to introduce a IDLETIMER v2.
Pablo Neira Ayuso
maintainerQuoted 18 lines · Jan Engelhardt
Quoted 16 lines · Maciej Żenczykowski
Signed-off-by Maciej Żenczykowski <[email protected]>---
include/uapi/linux/netfilter/xt_IDLETIMER.h| 1 +1 file changed, 1 insertion(+)include/uapi/linux/netfilter/xt_IDLETIMER.h +1
-0@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */ __u8 timer_type;
Quoted 7 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h struct idletimer_tg_info_v1 { · line 54
Quoted 2 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h +0
-0/* for kernel module internal use only */ This breaks the ABI for law-abiding Linux users (i.e. the GNU/Linux subgroup of it), which is equally terrible.
You will have to introduce a IDLETIMER v2.
IIRC, IDLETIMER v1 is in net-next, scheduled for 5.7-rc, there is no release for this code yet.
Right, this is not in 5.6 as it's only in net-next atm as it was only merged very recently.
I mentioned this in the commit message.
I'm not sure what you mean by code that uses this.
You can checkout aosp source and look there...
There's the kernel code (that's effectively already linked from the commit message), and the iptables userspace changes (
https://android.googlesource.com/platform/external/iptables/+/refs/heads/master/extensions/libxt_IDLETIMER.c#39
), and the netd C++/Java layer that uses iptables -j IDLETIMER --send_nl_msg 1 (
https://android.googlesource.com/platform/system/netd/+/refs/heads/master/server/IdletimerController.cpp#151
) and the resulting notifications parsing (can't easily find it atm).
If you mean by code that uses this patch... that's impossible as this patch doesn't implement a usable feature.
It just moves the offset.
Could you clarify what you're asking for?
Quoted 30 lines · Pablo Neira Ayuso
Quoted 18 lines · Jan Engelhardt
Quoted 16 lines · Maciej Żenczykowski
Signed-off-by Maciej Żenczykowski <[email protected]>---
include/uapi/linux/netfilter/xt_IDLETIMER.h| 1 +1 file changed, 1 insertion(+)include/uapi/linux/netfilter/xt_IDLETIMER.h +1
-0@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */ __u8 timer_type;Quoted 7 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h struct idletimer_tg_info_v1 {
Quoted 2 lines
include/uapi/linux/netfilter/xt_IDLETIMER.h +0
-1/* for kernel module internal use only */ --This breaks the ABI for law-abiding Linux users (i.e. the GNU/Linux subgroup of it), which is equally terrible.
You will have to introduce a IDLETIMER v2.
IIRC, IDLETIMER v1 is in net-next, scheduled for 5.7-rc, there is no release for this code yet.
Pablo Neira Ayuso
maintainerQuoted 9 lines · Maciej Żenczykowski
Right, this is not in 5.6 as it's only in net-next atm as it was only merged very recently.
I mentioned this in the commit message.I'm not sure what you mean by code that uses this.
You can checkout aosp source and look there...
There's the kernel code (that's effectively already linked from the commit message), and the iptables userspace changes (
https://android.googlesource.com/platform/external/iptables/+/refs/heads/master/extensions/libxt_IDLETIMER.c#39
OK, so this is field ised set in userspace.
Quoted 10 lines
), and the netd C++/Java layer that uses iptables -j IDLETIMER --send_nl_msg 1 (
https://android.googlesource.com/platform/system/netd/+/refs/heads/master/server/IdletimerController.cpp#151
) and the resulting notifications parsing (can't easily find it atm).If you mean by code that uses this patch... that's impossible as this patch doesn't implement a usable feature.
It just moves the offset.Could you clarify what you're asking for?
Maybe I'm misunderstanding. How is this field used in aosp?
I mean, if --send_nl_msg 1 is passed, how does the existing behaviour changes?
Right, so if you look at the android common kernel implementation.
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/net/netfilter/xt_IDLETIMER.c#201
(and in particular grep for send_nl_msg throughout the file) the core difference is the call to notify_netlink_uevent() at
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/net/netfilter/xt_IDLETIMER.c#101
this function is also defined in the same file, and it ends up calling
kobject_uevent_env(idletimer_tg_kobj, KOBJ_CHANGE, envp);
with INTERFACE, STATE, UID and TIME_NS metadata.
this is later parsed in netd C++:
https://android.googlesource.com/platform/system/netd/+/refs/heads/master/server/NetlinkHandler.cpp#208
which then via notifyInterfaceClassActivityChanged() I believe generates a binder call into the java network stack for further processing.
void NetlinkHandler::notifyInterfaceClassActivityChanged(int label, bool isActive,int64_t timestamp, int uid) { LOG_EVENT_FUNC(BINDER_RETRY, onInterfaceClassActivityChanged, isActive, label, timestamp, uid);
}
Ultimately the goal is to know WHO (uid - on Android it's 1 unix uid per {user,app} combination) generated traffic
activity, WHEN (timestamp) and WHERE (label ie. interface).
ie. To figure out whether an interface is idle or not and can be shutdown, or powersaved, etc.
WHO matters because some users (apps!) might be insufficiently important (lack of background data privs) to prevent
a network device powersavings/sleep/disconnect or to cause a wakeup (I think).
My understanding is the 'timer' thing (aka HARDIDLETIMER) was added because phones are almost constantly
entering suspend state, so we care about real clock time (CLOCK_BOOTTIME) rather than cpu activity time (CLOCK_MONOTONIC which doesn't include time the device is suspended).
I think this could probably be eventually switched to use the sysfs notification mechanism,
but we can't change that on any of the already shipped devices and it's kind of late even for devices shipping this year,
and that system would presumably need to be extended to support uids.
Quoted 29 lines · Pablo Neira Ayuso
Quoted 9 lines · Maciej Żenczykowski
Right, this is not in 5.6 as it's only in net-next atm as it was only merged very recently.
I mentioned this in the commit message.I'm not sure what you mean by code that uses this.
You can checkout aosp source and look there...
There's the kernel code (that's effectively already linked from the commit message), and the iptables userspace changes (
https://android.googlesource.com/platform/external/iptables/+/refs/heads/master/extensions/libxt_IDLETIMER.c#39OK, so this is field ised set in userspace.
Quoted 10 lines
), and the netd C++/Java layer that uses iptables -j IDLETIMER --send_nl_msg 1 (
https://android.googlesource.com/platform/system/netd/+/refs/heads/master/server/IdletimerController.cpp#151
) and the resulting notifications parsing (can't easily find it atm).If you mean by code that uses this patch... that's impossible as this patch doesn't implement a usable feature.
It just moves the offset.Could you clarify what you're asking for?
Maybe I'm misunderstanding. How is this field used in aosp?
I mean, if --send_nl_msg 1 is passed, how does the existing behaviour changes?
Quoted 2 lines · Maciej Żenczykowski
Right, this is not in 5.6 as it's only in net-next atm as it was only merged very recently.
Sorry. I read between the lines and had picked up "v0 is not present" and erroneously applied that statement to the standard kernel.
Pablo Neira Ayuso
maintainerHi Maciej,
Quoted 1 line · Maciej Żenczykowski
From: Maciej Żenczykowski <[email protected]>
[...]
Quoted 4 lines
We immediately notice that these two rev 1 structs are the same size and layout, and that while
timer_typeandsend_nl_msgare differently named and serve a different purpose, they're at the same offset.
I'm attaching a new version, including EOPNOTSUPP if the send_nl_msg field is set, it's the most basic handling I can think of until this option becomes useful.
Would you commit to send a patch for this new merge window to make it useful?
Thank you.
From f29c2a23ac49126886a2248e5948a8c62888e3dc Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Maciej=20=C5=BBenczykowski?= <[email protected]>
Date: Tue, 31 Mar 2020 09:35:59 -0700
Subject: [PATCH] netfilter: IDLETIMER target v1 - match Android layout
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Android has long had an extension to IDLETIMER to send netlink messages to userspace, see:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/include/uapi/linux/netfilter/xt_IDLETIMER.h#42
Note: this is idletimer target rev 1, there is no rev 0 in the Android common kernel sources, see registration at:
https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/net/netfilter/xt_IDLETIMER.c#483
When we compare that to upstream's new idletimer target rev 1:
https://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf-next.git/tree/include/uapi/linux/netfilter/xt_IDLETIMER.h#n46
We immediately notice that these two rev 1 structs are the same size and layout, and that while timer_type and send_nl_msg are differently named and serve a different purpose, they're at the same offset.
This makes them impossible to tell apart - and thus one cannot know in a mixed Android/vanilla environment whether one means timer_type or send_nl_msg.
Since this is iptables/netfilter uapi it introduces a problem between iptables (vanilla vs Android) userspace and kernel (vanilla vs Android) if the two don't match each other.
Additionally when at some point in the future Android picks up 5.7+ it's not at all clear how to resolve the resulting merge conflict.
Furthermore, since upgrading the kernel on old Android phones is pretty much impossible there does not seem to be an easy way out of this predicament.
The only thing I've been able to come up with is some super disgusting kernel version >= 5.7 check in the iptables binary to flip between different struct layouts.
By adding a dummy field to the vanilla Linux kernel header file we can force the two structs to be compatible with each other.
Long term I think I would like to deprecate send_nl_msg out of Android entirely, but I haven't quite been able to figure out exactly how we depend on it. It seems to be very similar to sysfs notifications but with some extra info.
Currently it's actually always enabled whenever Android uses the IDLETIMER target, so we could also probably entirely remove it from the uapi in favour of just always enabling it, but again we can't upgrade old kernels already in the field.
(Also note that this doesn't change the structure's size, as it is simply fitting into the pre-existing padding, and that since 5.7 hasn't been released yet, there's still time to make this uapi visible change)
--- include/uapi/linux/netfilter/xt_IDLETIMER.h | 1 + net/netfilter/xt_IDLETIMER.c | 3 +++ 2 files changed, 4 insertions(+)
include/uapi/linux/netfilter/xt_IDLETIMER.h +1 -0
@@ -48,6 +48,7 @@ struct idletimer_tg_info_v1 {char label[MAX_IDLETIMER_LABEL_SIZE ]; +__u8 send_nl_msg;/* unused: for compatibility with Android */ __u8 timer_type;/* for kernel module internal use only */
net/netfilter/xt_IDLETIMER.c +3 -0
@@ -346,6 +346,9 @@ static int idletimer_tg_checkentry_v1(const struct xt_tgchk_param *par)pr_debug ("checkentry targinfo%s\n" , info->label); +if (info->send_nl_msg) +return -EOPNOTSUPP ; + ret =idletimer_tg_helper ((struct idletimer_tg_info *)info);if (ret <0 ) {
Signature · 1 line
2.11.0
Maciej Żenczykowski
Reviewed-byQuoted 8 lines
Hi Maciej,
I'm attaching a new version, including EOPNOTSUPP if the
send_nl_msgfield is set, it's the most basic handling I can think of until this option becomes useful.Would you commit to send a patch for this new merge window to make it useful?
Yes, I can do that.
Thank you.
Pablo Neira Ayuso
maintainerQuoted 10 lines · Maciej Żenczykowski
Quoted 8 lines
Hi Maciej,
I'm attaching a new version, including EOPNOTSUPP if the
send_nl_msgfield is set, it's the most basic handling I can think of until this option becomes useful.Would you commit to send a patch for this new merge window to make it useful?
Yes, I can do that.
Thank you.
Patch is applied to nf.git