mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH net-next v3] mctp i2c: add sysfs control for mux child netdevs
@ 2026-11-18 10:02 Kenny Hsieh
  2026-11-21 23:02 ` netdev-bot+sashiko
  0 siblings, 1 reply; 2+ messages in thread
From: Kenny Hsieh @ 2026-11-18 10:02 UTC (permalink / raw)
  To: netdev, Jeremy Kerr, Matt Johnston, Andi Shyti, Peter Rosin
  Cc: Andrew Lunn, David S . Miller, Eric Dumazet, Jakub Kicinski,
	Paolo Abeni, linux-i2c, linux-kernel

An MCTP netdev is created for an I2C bus, root or mux channel, that
carries the "mctp-controller" devicetree property. A mux created at
runtime through the I2C sysfs interface has no devicetree node, and
neither do its channels, so they can never get a netdev.

This matters when the mux topology is only known at runtime. A BMC
that ships one firmware image across several chassis cannot describe
every chassis's mux tree in its devicetree, so userspace creates the
muxes it finds over I2C sysfs instead.

Add two write-only attributes to the mctp-i2c-controller device:

  echo i2c-30 > /sys/bus/i2c/devices/9-1010/add_mctp_controller
  echo i2c-30 > /sys/bus/i2c/devices/9-1010/remove_mctp_controller

Writing the name of a bus in the controller's mux tree creates or
removes the netdev for that bus.

The attributes belong to the driver, through dev_groups, so they exist
only while the controller is bound. Netdevs added through them are
removed on unbind and are not restored on rebind. The devicetree
property keeps working, and both can be used on the same mux tree.
mctp_i2c_add_netdev() now skips an adapter that already has a netdev,
so a probe walk or an add notifier that reaches an adapter a second
time no longer fails on the duplicate mctpi2c name.

The netdev is created in the writer's network namespace, and a write
requires CAP_NET_ADMIN there, as for a netdev created over rtnetlink.

Tested on an AST2700 BMC with a two-level mux tree: add and remove on
a channel without the devicetree property, unbind and bind of the
controller, and a loop of both while MCTP traffic runs on the channel.

Signed-off-by: Kenny Hsieh <kenny_hsieh@jabil.com>
---
v3:
  - Moved the attributes to the mctp-i2c-controller device, as Matt
    suggested. The driver core handles their lifetime now, so the
    tracking list and the races Sashiko found with it are gone.
  - Repeated add returns -EEXIST, remove with no netdev returns -ENODEV,
    per Matt.
  - Also remove on BUS_NOTIFY_REMOVED_DEVICE, so a write racing an
    adapter delete can't leave a netdev behind.
  - Dropped the stored netns, since get_net() kept it alive while the
    controller was bound. The netdev goes in the writer's netns now, as
    with rtnetlink, so there's no same-netns check anymore.
  - Dropped the midev = NULL change, the double free is fixed separately:
    https://lore.kernel.org/netdev/20261118090515.3390034-1-kenny_hsieh@jabil.com/
  - Rewrote the ABI doc, rebased on current net-next.

v2: https://lore.kernel.org/netdev/20261015081341.1262695-1-kenny_hsieh@jabil.com/
v1: https://lore.kernel.org/netdev/20261002073156.759005-1-kenny_hsieh@jabil.com/

 .../ABI/testing/sysfs-bus-i2c-devices-mctp    |  24 ++++
 MAINTAINERS                                   |   1 +
 drivers/net/mctp/mctp-i2c.c                   | 110 ++++++++++++++++--
 3 files changed, 124 insertions(+), 11 deletions(-)
 create mode 100644 Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp

diff --git a/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp b/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp
new file mode 100644
index 0000000000000..4ee96254f57a4
--- /dev/null
+++ b/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp
@@ -0,0 +1,24 @@
+What:		/sys/bus/i2c/devices/<bus>-<addr>/add_mctp_controller
+What:		/sys/bus/i2c/devices/<bus>-<addr>/remove_mctp_controller
+Date:		December 2026
+KernelVersion:	7.5
+Contact:	Kenny Hsieh <kenny_hsieh@jabil.com>
+Description:
+		Write-only, on each bound mctp-i2c-controller device.
+
+		Writing an I2C bus name such as "i2c-30" to add_mctp_controller
+		creates an MCTP network device for that bus, and writing it to
+		remove_mctp_controller removes it. The bus must be the
+		controller's root bus or a mux child below it. Adding a bus
+		that already has a network device fails with EEXIST, and
+		removing one that has none fails with ENODEV.
+
+		This covers busses instantiated at runtime, which have no
+		devicetree node for the "mctp-controller" property. Network
+		devices added here are removed on unbind and not restored on
+		rebind.
+
+		Example::
+
+		  # echo i2c-30 > /sys/bus/i2c/devices/9-1010/add_mctp_controller
+		  # echo i2c-30 > /sys/bus/i2c/devices/9-1010/remove_mctp_controller
diff --git a/MAINTAINERS b/MAINTAINERS
index 7d8b59ae6ea6c..ca07689318285 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -16093,6 +16093,7 @@ M:	Jeremy Kerr <jk@codeconstruct.com.au>
 M:	Matt Johnston <matt@codeconstruct.com.au>
 L:	netdev@vger.kernel.org
 S:	Maintained
+F:	Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp
 F:	Documentation/networking/mctp.rst
 F:	drivers/net/mctp/
 F:	include/linux/usb/mctp-usb.h
diff --git a/drivers/net/mctp/mctp-i2c.c b/drivers/net/mctp/mctp-i2c.c
index b02476d338bbf..61dd4a9e72545 100644
--- a/drivers/net/mctp/mctp-i2c.c
+++ b/drivers/net/mctp/mctp-i2c.c
@@ -857,6 +857,24 @@ static int mctp_i2c_ndo_open(struct net_device *dev)
 	return 0;
 }
 
+static struct mctp_i2c_dev *mctp_i2c_find_dev(struct mctp_i2c_client *mcli,
+					      struct i2c_adapter *adap)
+{
+	struct mctp_i2c_dev *midev = NULL, *m = NULL;
+	unsigned long flags;
+
+	spin_lock_irqsave(&mcli->sel_lock, flags);
+	/* List size is limited by number of MCTP netdevs on a single hardware bus */
+	list_for_each_entry(m, &mcli->devs, list)
+		if (m->adapter == adap) {
+			midev = m;
+			break;
+		}
+	spin_unlock_irqrestore(&mcli->sel_lock, flags);
+
+	return midev;
+}
+
 static int mctp_i2c_add_netdev(struct mctp_i2c_client *mcli,
 			       struct i2c_adapter *adap)
 {
@@ -876,6 +894,9 @@ static int mctp_i2c_add_netdev(struct mctp_i2c_client *mcli,
 	}
 
 	WARN_ON(!mutex_is_locked(&driver_clients_lock));
+	if (mctp_i2c_find_dev(mcli, adap))
+		return 0;
+
 	snprintf(namebuf, sizeof(namebuf), "mctpi2c%d", adap->nr);
 	ndev = alloc_netdev(sizeof(*midev), namebuf, NET_NAME_ENUM, mctp_i2c_net_setup);
 	if (!ndev) {
@@ -920,19 +941,10 @@ static int mctp_i2c_add_netdev(struct mctp_i2c_client *mcli,
 static void mctp_i2c_remove_netdev(struct mctp_i2c_client *mcli,
 				   struct i2c_adapter *adap)
 {
-	struct mctp_i2c_dev *midev = NULL, *m = NULL;
-	unsigned long flags;
+	struct mctp_i2c_dev *midev;
 
 	WARN_ON(!mutex_is_locked(&driver_clients_lock));
-	spin_lock_irqsave(&mcli->sel_lock, flags);
-	/* List size is limited by number of MCTP netdevs on a single hardware bus */
-	list_for_each_entry(m, &mcli->devs, list)
-		if (m->adapter == adap) {
-			midev = m;
-			break;
-		}
-	spin_unlock_irqrestore(&mcli->sel_lock, flags);
-
+	midev = mctp_i2c_find_dev(mcli, adap);
 	if (midev)
 		mctp_i2c_unregister(midev);
 }
@@ -1077,6 +1089,79 @@ static void mctp_i2c_remove(struct i2c_client *client)
 	mutex_unlock(&driver_clients_lock);
 }
 
+/* Returns the named adapter in mcli's mux tree, with a reference held */
+static struct i2c_adapter *mctp_i2c_find_named(struct mctp_i2c_client *mcli,
+					       const char *name)
+{
+	struct i2c_adapter *adap = NULL, *root = NULL;
+	struct device *dev;
+
+	dev = bus_find_device_by_name(&i2c_bus_type, NULL, name);
+	if (!dev)
+		return ERR_PTR(-ENODEV);
+
+	adap = mctp_i2c_get_adapter(dev, &root);
+	if (!adap || root != mcli->client->adapter) {
+		put_device(dev);
+		return ERR_PTR(-EINVAL);
+	}
+	return adap;
+}
+
+static ssize_t mctp_i2c_controller_write(struct device *dev, const char *buf,
+					 size_t count, bool add)
+{
+	struct mctp_i2c_client *mcli = i2c_get_clientdata(to_i2c_client(dev));
+	struct i2c_adapter *adap;
+	int rc = 0;
+
+	if (!ns_capable(current->nsproxy->net_ns->user_ns, CAP_NET_ADMIN))
+		return -EPERM;
+
+	mutex_lock(&driver_clients_lock);
+	adap = mctp_i2c_find_named(mcli, buf);
+	if (IS_ERR(adap)) {
+		rc = PTR_ERR(adap);
+		goto out;
+	}
+
+	if (add && mctp_i2c_find_dev(mcli, adap))
+		rc = -EEXIST;
+	else if (add)
+		rc = mctp_i2c_add_netdev(mcli, adap);
+	else if (!mctp_i2c_find_dev(mcli, adap))
+		rc = -ENODEV;
+	else
+		mctp_i2c_remove_netdev(mcli, adap);
+	put_device(&adap->dev);
+out:
+	mutex_unlock(&driver_clients_lock);
+	return rc < 0 ? rc : count;
+}
+
+static ssize_t add_mctp_controller_store(struct device *dev,
+					 struct device_attribute *attr,
+					 const char *buf, size_t count)
+{
+	return mctp_i2c_controller_write(dev, buf, count, true);
+}
+static DEVICE_ATTR_WO(add_mctp_controller);
+
+static ssize_t remove_mctp_controller_store(struct device *dev,
+					    struct device_attribute *attr,
+					    const char *buf, size_t count)
+{
+	return mctp_i2c_controller_write(dev, buf, count, false);
+}
+static DEVICE_ATTR_WO(remove_mctp_controller);
+
+static struct attribute *mctp_i2c_attrs[] = {
+	&dev_attr_add_mctp_controller.attr,
+	&dev_attr_remove_mctp_controller.attr,
+	NULL,
+};
+ATTRIBUTE_GROUPS(mctp_i2c);
+
 /* We look for a 'mctp-controller' property on I2C busses as they are
  * added/deleted, creating/removing netdevs as required.
  */
@@ -1090,6 +1175,8 @@ static int mctp_i2c_notifier_call(struct notifier_block *nb,
 		mctp_i2c_notify_add(dev);
 		break;
 	case BUS_NOTIFY_DEL_DEVICE:
+	/* A sysfs write can add a netdev after DEL_DEVICE */
+	case BUS_NOTIFY_REMOVED_DEVICE:
 		mctp_i2c_notify_del(dev);
 		break;
 	}
@@ -1116,6 +1203,7 @@ static struct i2c_driver mctp_i2c_driver = {
 	.driver = {
 		.name = "mctp-i2c-interface",
 		.of_match_table = mctp_i2c_of_match,
+		.dev_groups = mctp_i2c_groups,
 	},
 	.probe = mctp_i2c_probe,
 	.remove = mctp_i2c_remove,
-- 
2.50.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH net-next v3] mctp i2c: add sysfs control for mux child netdevs
  2026-11-18 10:02 [PATCH net-next v3] mctp i2c: add sysfs control for mux child netdevs Kenny Hsieh
@ 2026-11-21 23:02 ` netdev-bot+sashiko
  0 siblings, 0 replies; 2+ messages in thread
From: netdev-bot+sashiko @ 2026-11-21 23:02 UTC (permalink / raw)
  To: kenny_hsieh
  Cc: netdev, jk, matt, andi.shyti, peda, andrew+netdev, davem,
	edumazet, kuba, pabeni, linux-i2c, linux-kernel

Thank you for your contribution! Sashiko AI review found 3 potential
issue(s) to consider.

Critical: 0 · High: 0 (1) · Medium: 1 · Low: 2

If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).

- [Medium] mctp_i2c_controller_write() (drivers/net/mctp/mctp-i2c.c)
  checks permission with ns_capable(current->nsproxy->net_ns->user_ns,…
- [Low] Two inaccuracies in the commit message.
- [Low] Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp documents
  only two errors: EEXIST (adding a bus that already has a netdev) and…

Pre-existing issues:
- [High] This problem wasn't introduced by this patch, but the patch
  makes it reachable from userspace.

--

mctp i2c: add sysfs control for mux child netdevs

This patch adds write-only add_mctp_controller and remove_mctp_controller
attributes to the mctp-i2c-controller device. Writing a bus name to them
creates or removes an MCTP netdev for that bus in the controller's mux
tree, so muxes instantiated at runtime without a devicetree node can still
get a netdev. It also makes mctp_i2c_add_netdev() skip adapters that
already have a netdev, and it now handles BUS_NOTIFY_REMOVED_DEVICE in the
bus notifier.

> diff --git a/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp b/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp
> new file mode 100644
> index 0000000000000..4ee96254f57a4
> --- /dev/null
> +++ b/Documentation/ABI/testing/sysfs-bus-i2c-devices-mctp
> @@ -0,0 +1,24 @@
> +What:		/sys/bus/i2c/devices/<bus>-<addr>/add_mctp_controller
> +What:		/sys/bus/i2c/devices/<bus>-<addr>/remove_mctp_controller

[ ... ]

> +		Writing an I2C bus name such as "i2c-30" to add_mctp_controller
> +		creates an MCTP network device for that bus, and writing it to
> +		remove_mctp_controller removes it. The bus must be the
> +		controller's root bus or a mux child below it. Adding a bus
> +		that already has a network device fails with EEXIST, and
> +		removing one that has none fails with ENODEV.

[Severity: Low]
Should this ABI description list the other errors the store can return?

mctp_i2c_find_named() returns -ENODEV when bus_find_device_by_name() finds
nothing, on both add and remove:

	dev = bus_find_device_by_name(&i2c_bus_type, NULL, name);
	if (!dev)
		return ERR_PTR(-ENODEV);

On remove_mctp_controller, ENODEV can therefore mean "no such bus" or
"bus exists but has no netdev", and userspace can't tell which.

It also returns -EINVAL when the name is an i2c client rather than an
adapter, or an adapter outside this controller's mux tree.
mctp_i2c_controller_write() returns -EPERM when the CAP_NET_ADMIN check
fails.

The text also doesn't say that a CAP_NET_ADMIN check applies, or that the
netdev is created in the writer's network namespace.

[ ... ]

> diff --git a/drivers/net/mctp/mctp-i2c.c b/drivers/net/mctp/mctp-i2c.c
> index b02476d338bbf..61dd4a9e72545 100644
> --- a/drivers/net/mctp/mctp-i2c.c
> +++ b/drivers/net/mctp/mctp-i2c.c

[ ... ]

> @@ -876,6 +894,9 @@ static int mctp_i2c_add_netdev(struct mctp_i2c_client *mcli,
>  	}
>  
>  	WARN_ON(!mutex_is_locked(&driver_clients_lock));
> +	if (mctp_i2c_find_dev(mcli, adap))
> +		return 0;
> +

[Severity: Low]
The commit message gives this reason for the skip:

  mctp_i2c_add_netdev() now skips an adapter that already has a netdev,
  so a probe walk or an add notifier that reaches an adapter a second
  time no longer fails on the duplicate mctpi2c name.

Can a probe walk actually reach an adapter twice? mctp_i2c_probe()
allocates a new mcli with an empty devs list. It holds driver_clients_lock
for the whole i2c_for_each_dev() walk, and bus_for_each_dev() visits each
device only once. The dev_groups attributes are created only after probe
returns, so a sysfs add can't come first either.

The duplicates that can happen seem to be on the notifier side:

- an adapter is already on the bus klist, so the probe walk picks it up,
  and then BUS_NOTIFY_ADD_DEVICE fires for it;
- a sysfs add lands before the ADD_DEVICE notifier.

Could the commit message describe those cases instead?

The commit message body also doesn't mention that mctp_i2c_notifier_call()
now runs mctp_i2c_notify_del() on BUS_NOTIFY_REMOVED_DEVICE. That reason
appears only in the v3 notes below the cut line and in a code comment. The
v3 notes are dropped when the patch is applied.

>  	snprintf(namebuf, sizeof(namebuf), "mctpi2c%d", adap->nr);
>  	ndev = alloc_netdev(sizeof(*midev), namebuf, NET_NAME_ENUM, mctp_i2c_net_setup);
>  	if (!ndev) {

[ ... ]

> +static ssize_t mctp_i2c_controller_write(struct device *dev, const char *buf,
> +					 size_t count, bool add)
> +{
> +	struct mctp_i2c_client *mcli = i2c_get_clientdata(to_i2c_client(dev));
> +	struct i2c_adapter *adap;
> +	int rc = 0;
> +
> +	if (!ns_capable(current->nsproxy->net_ns->user_ns, CAP_NET_ADMIN))
> +		return -EPERM;

[Severity: Medium]
Is a check against the writer's own netns user namespace strong enough
here?

Any task can do unshare(CLONE_NEWUSER | CLONE_NEWNET) and then hold
CAP_NET_ADMIN in that namespace. So this check passes for anyone who can
open the attribute, and the root-owned 0200 file mode is the only real
gate. A privileged process could pass an open fd into a user-namespaced
context, or an admin could chown the attribute to a delegated user.

On add, mctp_i2c_add_netdev() does:

	dev_net_set(ndev, current->nsproxy->net_ns);

This puts the netdev for a physical I2C bus directly into a netns the
writer owns. Moving a physical netdev out of init_net over rtnetlink needs
CAP_NET_ADMIN in the source netns as well as the destination.

On remove, the branch below never checks dev_net(midev->ndev):

	else if (!mctp_i2c_find_dev(mcli, adap))
		rc = -ENODEV;
	else
		mctp_i2c_remove_netdev(mcli, adap);

A writer that has CAP_NET_ADMIN only in its own namespace can therefore
unregister an MCTP netdev in init_net. That includes one created at probe
time from the devicetree "mctp-controller" property. rtnetlink doesn't
allow that.

The commit message says "a write requires CAP_NET_ADMIN there, as for a
netdev created over rtnetlink". That comparison seems to hold only for
creating virtual devices.

Would capable(CAP_NET_ADMIN) fit better here? mctp_serial_open() in
drivers/net/mctp/mctp-serial.c already uses that check.

> +
> +	mutex_lock(&driver_clients_lock);
> +	adap = mctp_i2c_find_named(mcli, buf);
> +	if (IS_ERR(adap)) {
> +		rc = PTR_ERR(adap);
> +		goto out;
> +	}
> +
> +	if (add && mctp_i2c_find_dev(mcli, adap))
> +		rc = -EEXIST;
> +	else if (add)
> +		rc = mctp_i2c_add_netdev(mcli, adap);

[Severity: High]
This bug predates this patch, but the new store makes it reachable
from userspace. Can a failed registration free midev twice on this path?

mctp_i2c_add_netdev() calls mctp_register_netdev(), which reaches
register_netdevice(). Several of its failure paths call ndo_uninit:

- netdev_check_ops failure
- pcpu stats allocation failure
- a NETDEV_POST_INIT notifier veto
- a netdev_register_kobject() failure such as ENOMEM
- a NETDEV_REGISTER notifier veto, via unregister_netdevice_queue()

The first four reach err_uninit:

net/core/dev.c:register_netdevice() {
    ...
err_uninit:
	if (dev->netdev_ops->ndo_uninit)
		dev->netdev_ops->ndo_uninit(dev);
    ...
}

For this driver that is mctp_i2c_ndo_uninit()->mctp_i2c_midev_free(). That
function does list_del(&midev->list) and drops the get_device() references
on the adapter and the client.

The error then comes back to mctp_i2c_add_netdev(), which frees midev a
second time:

drivers/net/mctp/mctp-i2c.c:mctp_i2c_add_netdev() {
    ...
err:
	if (midev)
		mctp_i2c_midev_free(midev);
	if (ndev)
		free_netdev(ndev);
    ...
}

The second list_del() runs on poisoned pointers. The adapter and client
refcounts are also dropped twice, which can free those devices while they
are still in use.

Before this patch, only probe and the bus notifier reached this path. Now
add_mctp_controller lets a writer trigger it repeatedly.

The v3 notes point to a separate fix for this double free:

https://lore.kernel.org/netdev/20261118090515.3390034-1-kenny_hsieh@jabil.com/

That fix isn't in this tree yet, and the err: path above is unchanged.
Should that fix land before this patch, or together with it?

> +	else if (!mctp_i2c_find_dev(mcli, adap))
> +		rc = -ENODEV;
> +	else
> +		mctp_i2c_remove_netdev(mcli, adap);
> +	put_device(&adap->dev);
> +out:
> +	mutex_unlock(&driver_clients_lock);
> +	return rc < 0 ? rc : count;
> +}

[ ... ]

-- 
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261118100231.3827609-1-kenny_hsieh%40jabil.com

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-21 23:02 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-18 10:02 [PATCH net-next v3] mctp i2c: add sysfs control for mux child netdevs Kenny Hsieh
2026-11-21 23:02 ` netdev-bot+sashiko

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®