This adds a Kconfig option to set the default behavior regarding tunnel
fallback devices.
For setups where the initial namespace should also not have these, the
only preexisting option is to use a kernel command line option which
needs to be passed to every kernel invocation, which can be inconvenient
in certain setups.
If a kernel is built for a specific environment this knob allows
disabling the compatibility behavior outright, without requiring any
additional actions.

Signed-off-by: Lorenz Brun <lorenz@monogon.tech>
---
 net/Kconfig                | 33 +++++++++++++++++++++++++++++++++
 net/core/sysctl_net_core.c |  2 +-
 2 files changed, 34 insertions(+), 1 deletion(-)

diff --git a/net/Kconfig b/net/Kconfig
index d27d0deac0bf..e4429a017e47 100644
--- a/net/Kconfig
+++ b/net/Kconfig
@@ -447,6 +447,39 @@ config LWTUNNEL_BPF
 	  Allows to run BPF programs as a nexthop action following a route
 	  lookup for incoming and outgoing packets.
 
+choice
+	prompt "Create fallback tunnel devices"
+	default FB_TUNNELS_DEFAULT_ALL
+	help
+	  Fallback tunnel devices predate the Netlink API for managing network
+	  devices in Linux and get created when the respective tunnel kernel module
+	  is loaded. With a modern userspace these are no longer used but for
+	  compatibility reasons the default is to keep them around as the kernel
+	  cannot know if a given userspace needs them.
+	  There is a sysctl (net.core.fb_tunnels_only_for_init_net) for changing
+	  this, but it cannot retroactively remove fallback tunnel devices created
+	  before it was changed.
+
+	  This knob provides the possibility to set this behavior in the kernel,
+	  making it work in all cases. Note that changing this value to anything
+	  other than the default will break compatibility with old userspace.
+
+	config FB_TUNNELS_DEFAULT_ALL
+		bool "In every namespace"
+
+	config FB_TUNNELS_DEFAULT_INITNS
+		bool "Only in the initial namespace"
+
+	config FB_TUNNELS_DEFAULT_NONE
+		bool "Never"
+endchoice
+
+config FB_TUNNELS_DEFAULT
+	int
+	default 0 if FB_TUNNELS_DEFAULT_ALL
+	default 1 if FB_TUNNELS_DEFAULT_INITNS
+	default 2 if FB_TUNNELS_DEFAULT_NONE
+
 config DST_CACHE
 	bool
 	default n
diff --git a/net/core/sysctl_net_core.c b/net/core/sysctl_net_core.c
index 86a2476678c4..d9a0b13ceb4a 100644
--- a/net/core/sysctl_net_core.c
+++ b/net/core/sysctl_net_core.c
@@ -37,7 +37,7 @@ static int min_mem_pcpu_rsv = SK_MEMORY_PCPU_RESERVE;
 
 static int net_msg_warn;	/* Unused, but still a sysctl */
 
-int sysctl_fb_tunnels_only_for_init_net __read_mostly = 0;
+int sysctl_fb_tunnels_only_for_init_net __read_mostly = CONFIG_FB_TUNNELS_DEFAULT;
 EXPORT_SYMBOL(sysctl_fb_tunnels_only_for_init_net);
 
 /* 0 - Keep current behavior:
-- 
2.44.1

On Wed,  9 Oct 2024 13:04:19 +0200 Lorenz Brun wrote:
> This adds a Kconfig option to set the default behavior regarding tunnel
> fallback devices.
> For setups where the initial namespace should also not have these, the
> only preexisting option is to use a kernel command line option which
> needs to be passed to every kernel invocation, which can be inconvenient
> in certain setups.
> If a kernel is built for a specific environment this knob allows
> disabling the compatibility behavior outright, without requiring any
> additional actions.

I don't hate the idea, but could you comment a bit more on practical
usability of such a change? Whatever software depends on the new
behavior will likely not work with distro kernels, right?

Did you consider kernel boot param?