Skip to contentSkip to conversation

RDMA: Make device names unique per net namespace

PATCH Superseded Jiri Pirko 2 participants 23 messages Jul 14 – Jul 16

Subsystems (10)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jiri Pirko

author
#0 lore raw

v2 cover RDMA: Make device names unique per net namespace

From: Jiri Pirko <[email protected]>

RDMA device names are unique system-wide today:
__ib_device_get_by_name() checks a requested name against every registered device regardless of the network namespace it lives in. A device in one network namespace therefore cannot use a name already taken in another, even in exclusive netns mode (netns_mode=0) where the two are otherwise isolated. Net devices have no such restriction - their names only need to be unique within a network namespace.

This series makes RDMA device names unique per network namespace, matching net device semantics, and adapts the users that assumed system-wide unique names.

Scoping reuses the existing rdma_dev_access_netns() predicate, so behavior only changes in exclusive mode:
- shared mode (default): names stay unique system-wide, no change;
- exclusive mode: names only need to be unique within a namespace;
- CONFIG_NET_NS=n: everything is init_net, names stay system-wide unique.

There are two users that cannot be made per-namespace and are documented as known limitations instead of changed:
- the rdma_cm configfs tree: configfs has no network namespace

    support, so it cannot represent two same-named devices;

- SELinux ibendport labelling: endports are labelled by (device name, port) from a global policy; distinguishing same-named devices would need net namespace support in the SELinux policy language and tooling.

Tested with the new rxe_netns_names kselftest added in the last patch.

Jiri Pirko (14):
RDMA/core: Pass the net namespace to the device name lookups RDMA/core: Handle device name conflicts when changing net namespace RDMA/core: Support renaming a device when changing its net namespace RDMA/nldev: Report net namespace move errors through extack RDMA/nldev: Allow setting the device name while changing net namespace net/smc: Look up the pnetid ib device within the net namespace RDMA/srp: Make the SRP sysfs class net namespace aware RDMA/cgroup: Disambiguate devices across net namespaces RDMA/cma: Document that CM configfs cannot be net namespace scoped RDMA/core: Document the SELinux ibendport net namespace limitation RDMA/core: Make device names unique per net namespace RDMA/rxe: Allow queue VMAs to outlive ucontexts
RDMA/rxe: Implement disassociate_ucontext callback RDMA/selftests: Add rxe_netns_names test

 Documentation/ABI/testing/configfs-rdma_cm    |   4 +
 Documentation/admin-guide/cgroup-v1/rdma.rst  |   8 +
 Documentation/admin-guide/cgroup-v2.rst       |  15 +-
 drivers/infiniband/core/cgroup.c              |   1 +
 drivers/infiniband/core/cma_configfs.c        |   4 +
 drivers/infiniband/core/core_priv.h           |   3 +-
 drivers/infiniband/core/device.c              | 245 ++++++++++++---
 drivers/infiniband/core/nldev.c               |  26 +-
 drivers/infiniband/core/security.c            |   6 +
 drivers/infiniband/sw/rxe/rxe_mmap.c          |  35 +--
 drivers/infiniband/sw/rxe/rxe_verbs.c         |   5 +
 drivers/infiniband/ulp/srp/ib_srp.c           |  16 +-
 include/linux/cgroup_rdma.h                   |   1 +
 include/uapi/rdma/rdma_netlink.h              |   5 +-
 kernel/cgroup/rdma.c                          |  71 ++++-
 net/smc/smc_pnet.c                            |  20 +-
 tools/testing/selftests/rdma/Makefile         |   3 +-
 tools/testing/selftests/rdma/config           |   2 +
 .../testing/selftests/rdma/rxe_netns_names.sh | 282 ++++++++++++++++++
 19 files changed, 653 insertions(+), 99 deletions(-)
 create mode 100755 tools/testing/selftests/rdma/rxe_netns_names.sh
Signature · 1 line
2.54.0
author
#1 lore raw

PATCH v2 1/14 RDMA/core: Pass the net namespace to the device name lookups

From: Jiri Pirko <[email protected]>

Prepare for per-netns RDMA device names by passing the target net namespace through the name lookup and allocation helpers. Keep current global uniqueness behaviour.

---
 drivers/infiniband/core/device.c | 26 ++++++++++++++++++--------
 1 file changed, 18 insertions(+), 8 deletions(-)
drivers/infiniband/core/device.c +18 -8
index b8193e077a74..de610f52c9b2 100644
@@ -351,7 +351,8 @@ void ib_device_put(struct ib_device *device)
 }
 EXPORT_SYMBOL(ib_device_put);
 
-static struct ib_device *__ib_device_get_by_name(const char *name)
+static struct ib_device *__ib_device_get_by_name(const char *name,
+						 const struct net *net)
 {
 	struct ib_device *device;
 	unsigned long index;
@@ -395,7 +396,7 @@ int ib_device_rename(struct ib_device *ibdev, const char *name)
 		return 0;
 	}
 
-	if (__ib_device_get_by_name(name)) {
+	if (__ib_device_get_by_name(name, rdma_dev_net(ibdev))) {
 		up_write(&devices_rwsem);
 		return -EEXIST;
 	}
@@ -435,7 +436,8 @@ int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
 	return 0;
 }
 
-static int alloc_name(struct ib_device *ibdev, const char *name)
+/* Pick a free index for the '%d' style @name pattern. */
+static int alloc_name_id(struct net *net, const char *name)
 {
 	struct ib_device *device;
 	unsigned long index;
@@ -462,15 +464,22 @@ static int alloc_name(struct ib_device *ibdev, const char *name)
 	}
 
 	rc = ida_alloc(&inuse, GFP_KERNEL);
-	if (rc < 0)
-		goto out;
-
-	rc = dev_set_name(&ibdev->dev, name, rc);
 out:
 	ida_destroy(&inuse);
 	return rc;
 }
 
+static int alloc_name(struct ib_device *ibdev, const char *name)
+{
+	int id;
+
+	id = alloc_name_id(rdma_dev_net(ibdev), name);
+	if (id < 0)
+		return id;
+
+	return dev_set_name(&ibdev->dev, name, id);
+}
+
 static void ib_device_release(struct device *device)
 {
 	struct ib_device *dev = container_of(device, struct ib_device, dev);
@@ -1223,7 +1232,8 @@ static int assign_name(struct ib_device *device, const char *name)
 	if (ret)
 		goto out;
 
-	if (__ib_device_get_by_name(dev_name(&device->dev))) {
+	if (__ib_device_get_by_name(dev_name(&device->dev),
+				    rdma_dev_net(device))) {
 		ret = -ENFILE;
 		goto out;
 	}
Signature · 1 line
2.54.0
author
#2 lore raw

PATCH v2 2/14 RDMA/core: Handle device name conflicts when changing net namespace

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Prepare namespace moves for per-netns names. Check user-initiated moves for destination-name conflicts before disabling the device, keep same-netns moves as no-ops, and make teardown moves detach from the exiting namespace even if fallback naming fails.

---
v1->v2:
- fixed possible bogus race-triggered-WARN in rdma_dev_exit_net()

---
 drivers/infiniband/core/device.c | 155 ++++++++++++++++++++++++++-----
 drivers/infiniband/core/nldev.c  |   3 +
 2 files changed, 137 insertions(+), 21 deletions(-)
drivers/infiniband/core/device.c +134 -21
index de610f52c9b2..b55fb075d0ae 100644
@@ -268,7 +268,7 @@ static struct notifier_block ibdev_lsm_nb = {
 };
 
 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
-				 struct net *net);
+				 struct net *net, const char *fallback_pattern);
 
 /* Pointer to the RCU head at the start of the ib_port_data array */
 struct ib_port_data_rcu {
@@ -437,7 +437,8 @@ int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
 }
 
 /* Pick a free index for the '%d' style @name pattern. */
-static int alloc_name_id(struct net *net, const char *name)
+static int __alloc_name_id(struct net *net, const char *name,
+			   const struct ib_device *skip)
 {
 	struct ib_device *device;
 	unsigned long index;
@@ -450,6 +451,8 @@ static int alloc_name_id(struct net *net, const char *name)
 	xa_for_each (&devices, index, device) {
 		char buf[IB_DEVICE_NAME_MAX];
 
+		if (device == skip)
+			continue;
 		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 			continue;
 		if (i < 0 || i >= INT_MAX)
@@ -469,6 +472,11 @@ static int alloc_name_id(struct net *net, const char *name)
 	return rc;
 }
 
+static int alloc_name_id(struct net *net, const char *name)
+{
+	return __alloc_name_id(net, name, NULL);
+}
+
 static int alloc_name(struct ib_device *ibdev, const char *name)
 {
 	int id;
@@ -1160,8 +1168,17 @@ static void rdma_dev_exit_net(struct net *net)
 
 		/*
 		 * If the real device is in the NS then move it back to init.
+		 * Provide a fallback pattern so a name conflict in init_net
+		 * cannot make the teardown move fail.
 		 */
-		rdma_dev_change_netns(dev, net, &init_net);
+		if (net_eq(net, read_pnet(&dev->coredev.rdma_net))) {
+			ret = rdma_dev_change_netns(dev, net, &init_net,
+						    "ibdev%d");
+			if (ret && ret != -ENODEV)
+				WARN(1,
+				     "Failed to move RDMA device %s to init_net on netns exit: %d\n",
+				     dev_name(&dev->dev), ret);
+		}
 
 		put_device(&dev->dev);
 		down_read(&devices_rwsem);
@@ -1680,14 +1697,71 @@ void ib_unregister_device_queued(struct ib_device *ib_dev)
 }
 EXPORT_SYMBOL(ib_unregister_device_queued);
 
+static bool rdma_dev_name_in_netns(struct ib_device *skip, struct net *net,
+				   const char *name)
+{
+	struct ib_device *device;
+	unsigned long index;
+
+	lockdep_assert_held_write(&devices_rwsem);
+
+	xa_for_each(&devices, index, device)
+		if (device != skip &&
+		    !strcmp(name, dev_name(&device->dev)))
+			return true;
+
+	return false;
+}
+
+/*
+ * Choose the name @device should use in net namespace @net: keep the current
+ * name when it is free, otherwise use a trusted '%d' @fallback_pattern
+ * (namespace teardown) to pick a free index. The caller must hold the write
+ * side of devices_rwsem.
+ */
+static int rdma_dev_pick_netns_name(struct ib_device *device, struct net *net,
+				    const char *fallback_pattern,
+				    char *buf, size_t buf_len,
+				    const char **new_name)
+{
+	int id;
+
+	lockdep_assert_held_write(&devices_rwsem);
+
+	if (!rdma_dev_name_in_netns(device, net, dev_name(&device->dev))) {
+		*new_name = dev_name(&device->dev);
+		return 0;
+	}
+
+	if (!fallback_pattern)
+		return -EEXIST;
+
+	snprintf(buf, buf_len, "ibdev%u", device->index);
+	if (!rdma_dev_name_in_netns(device, net, buf)) {
+		*new_name = buf;
+		return 0;
+	}
+
+	id = __alloc_name_id(net, fallback_pattern, device);
+	if (id < 0)
+		return id;
+	snprintf(buf, buf_len, fallback_pattern, id);
+	*new_name = buf;
+	return 0;
+}
+
 /*
  * The caller must pass in a device that has the kref held and the refcount
  * released. If the device is in cur_net and still registered then it is moved
  * into net.
+ *
+ * Naming rules are handled by rdma_dev_pick_netns_name().
  */
 static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
-				 struct net *net)
+				 struct net *net, const char *fallback_pattern)
 {
+	char buf[IB_DEVICE_NAME_MAX];
+	const char *new_name;
 	int ret2 = -EINVAL;
 	int ret;
 
@@ -1704,30 +1778,63 @@ static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
 		goto out;
 	}
 
+	if (!fallback_pattern) {
+		/*
+		 * Reject a predictable name conflict before tearing anything
+		 * down, so a doomed user move does not disable a live device.
+		 */
+		down_write(&devices_rwsem);
+		ret = rdma_dev_pick_netns_name(device, net, fallback_pattern,
+					       buf, sizeof(buf), &new_name);
+		up_write(&devices_rwsem);
+		if (ret)
+			goto out;
+	}
+
 	kobject_uevent(&device->dev.kobj, KOBJ_REMOVE);
 	disable_device(device);
 
 	/*
-	 * At this point no one can be using the device, so it is safe to
-	 * change the namespace.
+	 * Recompute the destination name under the write side of devices_rwsem
+	 * now that the device is disabled, closing races with a concurrent
+	 * registration or rename, then publish the new namespace at the sysfs
+	 * level.
 	 */
-	write_pnet(&device->coredev.rdma_net, net);
+	down_write(&devices_rwsem);
+	ret = rdma_dev_pick_netns_name(device, net, fallback_pattern, buf,
+				       sizeof(buf), &new_name);
+	if (ret) {
+		if (fallback_pattern) {
+			WARN(1,
+			     "%s: failed to pick device name during namespace teardown: %d\n",
+			     __func__, ret);
+			write_pnet(&device->coredev.rdma_net, net);
+			ret = 0;
+		}
+		goto rename_done;
+	}
 
-	down_read(&devices_rwsem);
-	/*
-	 * Currently rdma devices are system wide unique. So the device name
-	 * is guaranteed free in the new namespace. Publish the new namespace
-	 * at the sysfs level.
-	 */
-	ret = device_rename(&device->dev, dev_name(&device->dev));
-	up_read(&devices_rwsem);
+	write_pnet(&device->coredev.rdma_net, net);
+	ret = device_rename(&device->dev, new_name);
 	if (ret) {
-		dev_warn(&device->dev,
-			 "%s: Couldn't rename device after namespace change\n",
-			 __func__);
-		/* Try and put things back and re-enable the device */
-		write_pnet(&device->coredev.rdma_net, cur_net);
+		if (fallback_pattern) {
+			WARN(1,
+			     "%s: failed to rename device during namespace teardown: %d\n",
+			     __func__, ret);
+			ret = 0;
+		} else {
+			dev_warn(&device->dev,
+				 "%s: Couldn't rename device after namespace change\n",
+				 __func__);
+			/* Try and put things back and re-enable the device */
+			write_pnet(&device->coredev.rdma_net, cur_net);
+		}
+	} else {
+		strscpy(device->name, dev_name(&device->dev),
+			IB_DEVICE_NAME_MAX);
 	}
+rename_done:
+	up_write(&devices_rwsem);
 
 	ret2 = enable_device_and_get(device);
 	if (ret2) {
@@ -1766,6 +1873,12 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 		goto ns_err;
 	}
 
+	/* Moving a device to the namespace it already lives in is a no-op. */
+	if (net_eq(net, read_pnet(&dev->coredev.rdma_net))) {
+		ret = 0;
+		goto ns_err;
+	}
+
 	/*
 	 * All the ib_clients, including uverbs, are reset when the namespace is
 	 * changed and this cannot be blocked waiting for userspace to do
@@ -1778,7 +1891,7 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 
 	get_device(&dev->dev);
 	ib_device_put(dev);
-	ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net);
+	ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net, NULL);
 	put_device(&dev->dev);
 
 	put_net(net);
drivers/infiniband/core/nldev.c +3 -0
index aae4f3f6bcba..32b6c4d68ca0 100644
@@ -1196,6 +1196,9 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 
 		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
 		err = ib_device_set_netns_put(skb, device, ns_fd);
+		if (err == -EEXIST)
+			NL_SET_ERR_MSG(extack,
+				       "Device name already exists in the target net namespace");
 		goto put_done;
 	}
 
Signature · 1 line
2.54.0
author
#3 lore raw

PATCH v2 3/14 RDMA/core: Support renaming a device when changing its net namespace

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Allow namespace moves to request a destination device name. Keep requested names on the same literal-name path as the existing RDMA rename operation, and keep teardown fallback naming on the trusted kernel-controlled path.

---
 drivers/infiniband/core/core_priv.h |  2 +-
 drivers/infiniband/core/device.c    | 48 ++++++++++++++++++++---------
 drivers/infiniband/core/nldev.c     |  2 +-
 3 files changed, 35 insertions(+), 17 deletions(-)
drivers/infiniband/core/core_priv.h +1 -1
index 19104c542b27..3bd5bb7135a3 100644
@@ -356,7 +356,7 @@ void ib_port_unregister_client_groups(struct ib_device *ibdev, u32 port_num,
 				     const struct attribute_group **groups);
 
 int ib_device_set_netns_put(struct sk_buff *skb,
-			    struct ib_device *dev, u32 ns_fd);
+			    struct ib_device *dev, u32 ns_fd, const char *name);
 
 int rdma_nl_net_init(struct rdma_dev_net *rnet);
 void rdma_nl_net_exit(struct rdma_dev_net *rnet);
drivers/infiniband/core/device.c +33 -15

111 lines not on this page · raw message

drivers/infiniband/core/nldev.c +1 -1
index 32b6c4d68ca0..77a758080148 100644
@@ -1195,7 +1195,7 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 		u32 ns_fd;
 
 		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
-		err = ib_device_set_netns_put(skb, device, ns_fd);
+		err = ib_device_set_netns_put(skb, device, ns_fd, NULL);
 		if (err == -EEXIST)
 			NL_SET_ERR_MSG(extack,
 				       "Device name already exists in the target net namespace");
Signature · 1 line
2.54.0
#4 lore raw

Re: [PATCH rdma-next v2 03/14] RDMA/core: Support renaming a device when changing its net namespace

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [High] Unreachable feature and silent failure in netlink handler for RDMA namespace moves --

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Unreachable feature and silent failure in netlink handler for RDMA namespace moves --

--- Patch [3]: [PATCH rdma-next v2 03/14] RDMA/core: Support renaming a device when changing its net namespace --- commit 4056da04c986796c81666352d7fa9e90d5591d4e
Author: Jiri Pirko <[email protected]>

RDMA/core: Support renaming a device when changing its net namespace

This commit introduces support for renaming an RDMA device when moving it to a new network namespace. It preserves the literal-name path for user requests and keeps the fallback naming on the teardown path.

Quoted 4 lines
drivers/infiniband/core/nldev.c +0 -0
index 32b6c4d68ca0a..77a7580801486 100644

[ ... ]

drivers/infiniband/core/nldev.c static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh, · line 1198

Quoted 6 lines
drivers/infiniband/core/nldev.c +1 -1
@@ -1195,7 +1195,7 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 		u32 ns_fd;
 
 		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
-		err = ib_device_set_netns_put(skb, device, ns_fd);
+		err = ib_device_set_netns_put(skb, device, ns_fd, NULL);

[Severity: High]
Does hardcoding NULL here make the newly introduced naming feature unreachable from user space?

If a user space application sends an RDMA_NLDEV_CMD_SET netlink message containing both RDMA_NLDEV_ATTR_DEV_NAME and RDMA_NLDEV_NET_NS_FD, nldev_set_doit() appears to hardcode NULL for the destination name instead of extracting and passing the requested name to ib_device_set_netns_put().

Additionally, if the RDMA_NLDEV_ATTR_DEV_NAME attribute is processed separately in nldev_set_doit():

    if (tb[RDMA_NLDEV_ATTR_DEV_NAME]) {
        ...
        err = ib_device_rename(device, name);
        goto done;
    }

is there a risk that passing both attributes will cause the namespace move to be silently skipped entirely because the rename block unconditionally jumps to the end of the function?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
author
#5 lore raw

Re: [PATCH rdma-next v2 03/14] RDMA/core: Support renaming a device when changing its net namespace

Quoted 22 lines · Wed, Jul 15, 2026 at 04:30:37PM +0200, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] Unreachable feature and silent failure in netlink handler for RDMA namespace moves --

--- Patch [3]: [PATCH rdma-next v2 03/14] RDMA/core: Support renaming a device when changing its net namespace --- commit 4056da04c986796c81666352d7fa9e90d5591d4e
Author: Jiri Pirko <[email protected]>

RDMA/core: Support renaming a device when changing its net namespace

This commit introduces support for renaming an RDMA device when moving it to a new network namespace. It preserves the literal-name path for user requests and keeps the fallback naming on the teardown path.

Quoted 4 lines
drivers/infiniband/core/nldev.c +0 -0
index 32b6c4d68ca0a..77a7580801486 100644

[ ... ]

Quoted 3 lines
drivers/infiniband/core/nldev.c +0 -0
@@ -1195,7 +1195,7 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 		u32 ns_fd;
 
Quoted 7 lines

drivers/infiniband/core/nldev.c static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh, · line 1198

Quoted 3 lines
drivers/infiniband/core/nldev.c +1 -1
 		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
-		err = ib_device_set_netns_put(skb, device, ns_fd);
+		err = ib_device_set_netns_put(skb, device, ns_fd, NULL);

[Severity: High]
Does hardcoding NULL here make the newly introduced naming feature unreachable from user space?

False positive. "RDMA/nldev: Allow setting the device name while changing net namespace" adds this.

author
#6 lore raw

PATCH v2 4/14 RDMA/nldev: Report net namespace move errors through extack

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Thread extack through the existing net namespace move helper and report the main failure reasons from the core path. Keep the existing move UAPI shape unchanged.

---
 drivers/infiniband/core/core_priv.h |  3 ++-
 drivers/infiniband/core/device.c    | 24 ++++++++++++++++++++++--
 drivers/infiniband/core/nldev.c     |  6 ++----
 3 files changed, 26 insertions(+), 7 deletions(-)
drivers/infiniband/core/core_priv.h +2 -1
index 3bd5bb7135a3..aaf330b0d333 100644
@@ -356,7 +356,8 @@ void ib_port_unregister_client_groups(struct ib_device *ibdev, u32 port_num,
 				     const struct attribute_group **groups);
 
 int ib_device_set_netns_put(struct sk_buff *skb,
-			    struct ib_device *dev, u32 ns_fd, const char *name);
+			    struct ib_device *dev, u32 ns_fd, const char *name,
+			    struct netlink_ext_ack *extack);
 
 int rdma_nl_net_init(struct rdma_dev_net *rnet);
 void rdma_nl_net_exit(struct rdma_dev_net *rnet);
drivers/infiniband/core/device.c +22 -2
index c0b6613dba4c..366bd8463c07 100644
@@ -1871,18 +1871,22 @@ static int rdma_dev_change_netns(struct ib_device *device, struct net *cur_net,
 }
 
 int ib_device_set_netns_put(struct sk_buff *skb,
-			    struct ib_device *dev, u32 ns_fd, const char *name)
+			    struct ib_device *dev, u32 ns_fd, const char *name,
+			    struct netlink_ext_ack *extack)
 {
 	struct net *net;
 	int ret;
 
 	net = get_net_ns_by_fd(ns_fd);
 	if (IS_ERR(net)) {
+		NL_SET_ERR_MSG(extack, "Invalid target net namespace fd");
 		ret = PTR_ERR(net);
 		goto net_err;
 	}
 
 	if (!netlink_ns_capable(skb, net->user_ns, CAP_NET_ADMIN)) {
+		NL_SET_ERR_MSG(extack,
+			       "Missing CAP_NET_ADMIN in the target net namespace");
 		ret = -EPERM;
 		goto ns_err;
 	}
@@ -1893,6 +1897,10 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 	 */
 	if (net_eq(net, read_pnet(&dev->coredev.rdma_net))) {
 		ret = name ? ib_device_rename(dev, name) : 0;
+
+		if (ret == -EEXIST)
+			NL_SET_ERR_MSG(extack,
+				       "Device name already exists in the target net namespace");
 		goto ns_err;
 	}
 
@@ -1901,7 +1909,16 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 	 * changed and this cannot be blocked waiting for userspace to do
 	 * something, so disassociation is mandatory.
 	 */
-	if (!dev->ops.disassociate_ucontext || ib_devices_shared_netns) {
+	if (ib_devices_shared_netns) {
+		NL_SET_ERR_MSG(extack,
+			       "Cannot change net namespace of RDMA device in shared netns mode");
+		ret = -EOPNOTSUPP;
+		goto ns_err;
+	}
+
+	if (!dev->ops.disassociate_ucontext) {
+		NL_SET_ERR_MSG(extack,
+			       "Device does not support namespace changes (no disassociate support)");
 		ret = -EOPNOTSUPP;
 		goto ns_err;
 	}
@@ -1911,6 +1928,9 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 	ret = rdma_dev_change_netns(dev, current->nsproxy->net_ns, net, name,
 				    NULL);
 	put_device(&dev->dev);
+	if (ret == -EEXIST)
+		NL_SET_ERR_MSG(extack,
+			       "Device name already exists in the target net namespace");
 
 	put_net(net);
 	return ret;
drivers/infiniband/core/nldev.c +2 -4
index 77a758080148..8648e95700bf 100644
@@ -1195,10 +1195,8 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 		u32 ns_fd;
 
 		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
-		err = ib_device_set_netns_put(skb, device, ns_fd, NULL);
-		if (err == -EEXIST)
-			NL_SET_ERR_MSG(extack,
-				       "Device name already exists in the target net namespace");
+		err = ib_device_set_netns_put(skb, device, ns_fd, NULL,
+					      extack);
 		goto put_done;
 	}
 
Signature · 1 line
2.54.0
author
#7 lore raw

PATCH v2 5/14 RDMA/nldev: Allow setting the device name while changing net namespace

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Accept RDMA_NLDEV_ATTR_DEV_NAME together with RDMA_NLDEV_NET_NS_FD so a netlink move can rename the device in the destination namespace. Keep the name semantics aligned with the existing RDMA rename path.

---
 drivers/infiniband/core/device.c |  6 ++++++
 drivers/infiniband/core/nldev.c  | 27 ++++++++++++++++++---------
 include/uapi/rdma/rdma_netlink.h |  5 ++++-
 3 files changed, 28 insertions(+), 10 deletions(-)
drivers/infiniband/core/device.c +6 -0
index 366bd8463c07..bb02640239d7 100644
@@ -1901,6 +1901,9 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 		if (ret == -EEXIST)
 			NL_SET_ERR_MSG(extack,
 				       "Device name already exists in the target net namespace");
+		else if (ret == -EINVAL && name)
+			NL_SET_ERR_MSG(extack,
+				       "Unable to use requested device name in the target net namespace");
 		goto ns_err;
 	}
 
@@ -1931,6 +1934,9 @@ int ib_device_set_netns_put(struct sk_buff *skb,
 	if (ret == -EEXIST)
 		NL_SET_ERR_MSG(extack,
 			       "Device name already exists in the target net namespace");
+	else if (ret == -EINVAL && name)
+		NL_SET_ERR_MSG(extack,
+			       "Unable to use requested device name in the target net namespace");
 
 	put_net(net);
 	return ret;
drivers/infiniband/core/nldev.c +18 -9
index 8648e95700bf..4efa387ec1be 100644
@@ -1178,6 +1178,24 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 	if (!device)
 		return -EINVAL;
 
+	if (tb[RDMA_NLDEV_NET_NS_FD]) {
+		char name[IB_DEVICE_NAME_MAX] = {};
+		u32 ns_fd;
+
+		if (tb[RDMA_NLDEV_ATTR_DEV_NAME]) {
+			nla_strscpy(name, tb[RDMA_NLDEV_ATTR_DEV_NAME],
+				    IB_DEVICE_NAME_MAX);
+			if (strlen(name) == 0) {
+				err = -EINVAL;
+				goto done;
+			}
+		}
+		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
+		err = ib_device_set_netns_put(skb, device, ns_fd,
+					      name[0] ? name : NULL, extack);
+		goto put_done;
+	}
+
 	if (tb[RDMA_NLDEV_ATTR_DEV_NAME]) {
 		char name[IB_DEVICE_NAME_MAX] = {};
 
@@ -1191,15 +1209,6 @@ static int nldev_set_doit(struct sk_buff *skb, struct nlmsghdr *nlh,
 		goto done;
 	}
 
-	if (tb[RDMA_NLDEV_NET_NS_FD]) {
-		u32 ns_fd;
-
-		ns_fd = nla_get_u32(tb[RDMA_NLDEV_NET_NS_FD]);
-		err = ib_device_set_netns_put(skb, device, ns_fd, NULL,
-					      extack);
-		goto put_done;
-	}
-
 	if (tb[RDMA_NLDEV_ATTR_DEV_DIM]) {
 		u8 use_dim;
 
include/uapi/rdma/rdma_netlink.h +4 -1
index 3af946ecbac3..ee11c3bbbae2 100644
@@ -516,7 +516,10 @@ enum rdma_nldev_attr {
 	RDMA_NLDEV_ATTR_DEV_PROTOCOL,		/* string */
 
 	/*
-	 * File descriptor handle of the net namespace object
+	 * File descriptor handle of the net namespace object. May be combined
+	 * with RDMA_NLDEV_ATTR_DEV_NAME (a literal device name) to also rename
+	 * the device in the destination namespace; the move fails with -EEXIST
+	 * if that name is already taken there.
 	 */
 	RDMA_NLDEV_NET_NS_FD,			/* u32 */
 	/*
Signature · 1 line
2.54.0
author
#8 lore raw

PATCH v2 6/14 net/smc: Look up the pnetid ib device within the net namespace

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Scope smc_pnet_find_ib() to the caller's net namespace so pnetid setup cannot bind to a same-named RDMA device from another namespace once names become per-netns.

---
 net/smc/smc_pnet.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)
net/smc/smc_pnet.c +14 -6
index 63e286e2dfaa..ff9c9c35cc2f 100644
@@ -304,13 +304,18 @@ static bool smc_pnetid_valid(const char *pnet_name, char *pnetid)
 	return true;
 }
 
-/* Find an infiniband device by a given name. The device might not exist. */
-static struct smc_ib_device *smc_pnet_find_ib(char *ib_name)
+/*
+ * Find an infiniband device by a given name, restricted to the devices
+ * accessible from @net. The device might not exist.
+ */
+static struct smc_ib_device *smc_pnet_find_ib(struct net *net, char *ib_name)
 {
 	struct smc_ib_device *ibdev;
 
 	mutex_lock(&smc_ib_devices.mutex);
 	list_for_each_entry(ibdev, &smc_ib_devices.list, list) {
+		if (!rdma_dev_access_netns(ibdev->ibdev, net))
+			continue;
 		if (!strncmp(ibdev->ibdev->name, ib_name,
 			     sizeof(ibdev->ibdev->name)) ||
 		    (ibdev->ibdev->dev.parent &&
@@ -408,8 +413,8 @@ static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 	return rc;
 }
 
-static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
-			   u8 ib_port, char *pnet_name)
+static int smc_pnet_add_ib(struct smc_pnettable *pnettable, struct net *net,
+			   char *ib_name, u8 ib_port, char *pnet_name)
 {
 	struct smc_pnetentry *tmp_pe, *new_pe;
 	struct smc_ib_device *ib_dev;
@@ -419,7 +424,7 @@ static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 	bool new_ibdev;
 
 	/* try to apply the pnetid to active devices */
-	ib_dev = smc_pnet_find_ib(ib_name);
+	ib_dev = smc_pnet_find_ib(net, ib_name);
 	if (ib_dev) {
 		ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 		if (ibdev_applied)
@@ -518,7 +523,7 @@ static int smc_pnet_enter(struct net *net, struct nlattr *tb[])
 			if (ibport < 1 || ibport > SMC_MAX_PORTS)
 				goto error;
 		}
-		rc = smc_pnet_add_ib(pnettable, string, ibport, pnet_name);
+		rc = smc_pnet_add_ib(pnettable, net, string, ibport, pnet_name);
 		if (!rc)
 			new_ibdev = true;
 		else if (rc != -EEXIST)
@@ -1170,6 +1175,9 @@ int smc_pnetid_by_table_ib(struct smc_ib_device *smcibdev, u8 ib_port)
 	struct smc_net *sn;
 	int rc = -ENOENT;
 
+	if (!rdma_dev_access_netns(smcibdev->ibdev, &init_net))
+		return -ENOENT;
+
 	/* get pnettable for init namespace */
 	sn = net_generic(&init_net, smc_net_id);
 	pnettable = &sn->pnettable;
Signature · 1 line
2.54.0
#9 lore raw

Re: [PATCH rdma-next v2 06/14] net/smc: Look up the pnetid ib device within the net namespace

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Returning unreferenced device pointers from smc_pnet_find_ib() and smc_pnet_find_smcd() causes Use-After-Free vulnerabilities when devices are concurrently removed. --

--- Patch [6]: [PATCH rdma-next v2 06/14] net/smc: Look up the pnetid ib device within the net namespace --- commit 4529cf06709604767797679d2d489bb89001f950

Author: Jiri Pirko <[email protected]>
Date:   Tue Jul 14 14:29:19 2026 +0000

net/smc: Look up the pnetid ib device within the net namespace

Scope smc_pnet_find_ib() to the caller's net namespace so pnetid setup cannot bind to a same-named RDMA device from another namespace once names become per-netns.

Quoted 16 lines
net/smc/smc_pnet.c +2 -2
index 63e286e2dfaa8..ff9c9c35cc2f5 100644
@@ -408,8 +413,8 @@ static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 	return rc;
 }
 
-static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
-			   u8 ib_port, char *pnet_name)
+static int smc_pnet_add_ib(struct smc_pnettable *pnettable, struct net *net,
+			   char *ib_name, u8 ib_port, char *pnet_name)
 {
 	struct smc_pnetentry *tmp_pe, *new_pe;
 	struct smc_ib_device *ib_dev;

net/smc/smc_pnet.c static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name, · line 427

Quoted 5 lines
net/smc/smc_pnet.c +1 -1
@@ -419,7 +424,7 @@ static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 	bool new_ibdev;
 
 	/* try to apply the pnetid to active devices */
-	ib_dev = smc_pnet_find_ib(ib_name);
+	ib_dev = smc_pnet_find_ib(net, ib_name);

[Severity: High]
This is a pre-existing issue, but can this lead to a use-after-free?

Looking at smc_pnet_find_ib(), it searches for an IB device while holding smc_ib_devices.mutex, finds one, drops the lock, and returns a raw, unreferenced ibdev pointer:

net/smc/smc_pnet.c:smc_pnet_find_ib() {
    ...
    mutex_unlock(&smc_ib_devices.mutex);
    return ibdev;
}

If an asynchronous device unregistration occurs immediately after the lock is dropped, smc_ib_remove_dev() could acquire the mutex, remove the device from the list, and free the smc_ib_device structure.

The caller (smc_pnet_add_ib) then dereferences the freed pointer locklessly (e.g., when accessing ib_dev->ibdev->name for logging) or passes it to smc_pnet_apply_ib(), which re-acquires the mutex and accesses fields of the potentially freed device pointer.

Should smc_pnet_find_ib() take a reference on the device before dropping the mutex?

Quoted 3 lines
net/smc/smc_pnet.c +0 -0
 	if (ib_dev) {
 		ibdev_applied = smc_pnet_apply_ib(ib_dev, ib_port, pnet_name);
 		if (ibdev_applied)
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=6
author
#10 lore raw

Re: [PATCH rdma-next v2 06/14] net/smc: Look up the pnetid ib device within the net namespace

Quoted 35 lines · Wed, Jul 15, 2026 at 04:30:42PM +0200, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Returning unreferenced device pointers from smc_pnet_find_ib() and smc_pnet_find_smcd() causes Use-After-Free vulnerabilities when devices are concurrently removed. --

--- Patch [6]: [PATCH rdma-next v2 06/14] net/smc: Look up the pnetid ib device within the net namespace --- commit 4529cf06709604767797679d2d489bb89001f950

Author: Jiri Pirko <[email protected]>
Date:   Tue Jul 14 14:29:19 2026 +0000

net/smc: Look up the pnetid ib device within the net namespace

Scope smc_pnet_find_ib() to the caller's net namespace so pnetid setup cannot bind to a same-named RDMA device from another namespace once names become per-netns.

Quoted 18 lines
net/smc/smc_pnet.c +2 -2
index 63e286e2dfaa8..ff9c9c35cc2f5 100644
@@ -408,8 +413,8 @@ static int smc_pnet_add_eth(struct smc_pnettable *pnettable, struct net *net,
 	return rc;
 }
 
-static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
-			   u8 ib_port, char *pnet_name)
+static int smc_pnet_add_ib(struct smc_pnettable *pnettable, struct net *net,
+			   char *ib_name, u8 ib_port, char *pnet_name)
 {
 	struct smc_pnetentry *tmp_pe, *new_pe;
 	struct smc_ib_device *ib_dev;
@@ -419,7 +424,7 @@ static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name,
 	bool new_ibdev;
 
Quoted 6 lines

net/smc/smc_pnet.c static int smc_pnet_add_ib(struct smc_pnettable *pnettable, char *ib_name, · line 427

Quoted 3 lines
net/smc/smc_pnet.c +1 -1
 	/* try to apply the pnetid to active devices */
-	ib_dev = smc_pnet_find_ib(ib_name);
+	ib_dev = smc_pnet_find_ib(net, ib_name);

[Severity: High]
This is a pre-existing issue, but can this lead to a use-after-free?

Unrelated to this patch/set. Ignoring.

author
#11 lore raw

PATCH v2 7/14 RDMA/srp: Make the SRP sysfs class net namespace aware

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Tag srp_class by the RDMA device's net namespace so SRP hosts derived from same-named RDMA devices can coexist across namespaces.

---
 drivers/infiniband/ulp/srp/ib_srp.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
drivers/infiniband/ulp/srp/ib_srp.c +15 -1
index 0caebbc2810f..2fc8e133c70f 100644
@@ -3189,10 +3189,24 @@ static struct attribute *srp_class_attrs[];
 
 ATTRIBUTE_GROUPS(srp_class);
 
+/*
+ * SRP hosts are named after their ib device, so tag the class by the ib
+ * device's net namespace.
+ */
+static const struct ns_common *srp_net_namespace(const struct device *dev)
+{
+	struct srp_host *host = container_of(dev, struct srp_host, dev);
+	struct net *net = rdma_dev_net(host->srp_dev->dev);
+
+	return net ? to_ns_common(net) : NULL;
+}
+
 static struct class srp_class = {
 	.name    = "infiniband_srp",
 	.dev_groups = srp_class_groups,
-	.dev_release = srp_release_dev
+	.dev_release = srp_release_dev,
+	.ns_type = &net_ns_type_operations,
+	.namespace = srp_net_namespace,
 };
 
 /**
Signature · 1 line
2.54.0
author
#12 lore raw

PATCH v2 8/14 RDMA/cgroup: Disambiguate devices across net namespaces

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

RDMA device names are unique only within a network namespace, but an RDMA cgroup can account resources for devices from multiple namespaces. Duplicate names therefore make cgroup output ambiguous and can cause limit writes to select the wrong device.

Use the system-wide RDMA device index to distinguish duplicate names while preserving the existing UAPI for unique names. Reject ambiguous name-only writes with -ENOTUNIQ and expose a complete device view to administrators.

---
v1->v2:
- replace caller-netns filtering with global visibility and index-based disambiguation

---
 Documentation/admin-guide/cgroup-v1/rdma.rst |  8 +++
 Documentation/admin-guide/cgroup-v2.rst      | 15 ++++-
 drivers/infiniband/core/cgroup.c             |  1 +
 include/linux/cgroup_rdma.h                  |  1 +
 kernel/cgroup/rdma.c                         | 71 ++++++++++++++++----
 5 files changed, 83 insertions(+), 13 deletions(-)
Documentation/admin-guide/cgroup-v1/rdma.rst +8 -0
index e69369b7252e..8d0c3a796ee3 100644
@@ -90,6 +90,13 @@ Following resources can be accounted by rdma controller.
   hca_object 	Maximum number of HCA Objects
   ==========    =============================
 
+RDMA devices from all network namespaces are listed. Each line starts with
+the device name. If more than one device has the same name, ``index=N``
+follows the name, where ``N`` is the system-wide RDMA device index, unique
+among registered devices. When configuring a limit, the index is optional
+for a globally unique name and required for a duplicate name. A write without
+the required index fails with ``-ENOTUNIQ``.
+
 2. Usage Examples
 =================
 
@@ -97,6 +104,7 @@ Following resources can be accounted by rdma controller.
 
 	echo mlx4_0 hca_handle=2 hca_object=2000 > /sys/fs/cgroup/rdma/1/rdma.max
 	echo ocrdma1 hca_handle=3 > /sys/fs/cgroup/rdma/2/rdma.max
+	echo "rxe0 index=5 hca_handle=2" > /sys/fs/cgroup/rdma/3/rdma.max
 
 (b) Query resource limit::
 
Documentation/admin-guide/cgroup-v2.rst +14 -1
index 993446ab66d0..df74d554d2cd 100644
@@ -2752,6 +2752,11 @@ RDMA
 The "rdma" controller regulates the distribution and accounting of
 RDMA resources.
 
+RDMA devices from all network namespaces are listed. Each line starts with
+the device name. If more than one device has the same name, ``index=N``
+follows the name, where ``N`` is the system-wide RDMA device index, unique
+among registered devices.
+
 RDMA Interface Files
 ~~~~~~~~~~~~~~~~~~~~
 
@@ -2760,7 +2765,11 @@ RDMA Interface Files
 	except root that describes current configured resource limit
 	for a RDMA/IB device.
 
-	Lines are keyed by device name and are not ordered.
+	Lines are keyed by device name and are not ordered. A write may
+	include ``index=N`` after the device name. The index is optional
+	when the name is globally unique. If multiple devices have that
+	name, the index is required and a write without it fails with
+	``-ENOTUNIQ``.
 	Each line contains space separated resource name and its configured
 	limit that can be distributed.
 
@@ -2776,6 +2785,10 @@ RDMA Interface Files
 	  mlx4_0 hca_handle=2 hca_object=2000
 	  ocrdma1 hca_handle=3 hca_object=max
 
+	For devices with duplicate names, select the device by index::
+
+	  echo "rxe0 index=5 hca_handle=2" > rdma.max
+
   rdma.current
 	A read-only file that describes current resource usage.
 	It exists for all the cgroup except root.
drivers/infiniband/core/cgroup.c +1 -0
index 1f037fe01450..8611b4e32cfb 100644
@@ -17,6 +17,7 @@
 void ib_device_register_rdmacg(struct ib_device *device)
 {
 	device->cg_device.name = device->name;
+	device->cg_device.index = device->index;
 	rdmacg_register_device(&device->cg_device);
 }
 
include/linux/cgroup_rdma.h +1 -0
index 404e746552ca..9a5c9ee728e7 100644
@@ -34,6 +34,7 @@ struct rdmacg_device {
 	struct list_head	dev_node;
 	struct list_head	rpools;
 	char			*name;
+	u32			index;
 };
 
 /*
kernel/cgroup/rdma.c +59 -12

137 lines not on this page · raw message

Signature · 1 line
2.54.0
author
#13 lore raw

PATCH v2 9/14 RDMA/cma: Document that CM configfs cannot be net namespace scoped

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Document the rdma_cm configfs limitation: configfs is global, so same-named RDMA devices in different net namespaces cannot both be represented there.

---
 Documentation/ABI/testing/configfs-rdma_cm | 4 ++++
 drivers/infiniband/core/cma_configfs.c     | 4 ++++
 2 files changed, 8 insertions(+)
Documentation/ABI/testing/configfs-rdma_cm +4 -0
index 74f9506f42e7..739f7b6a1259 100644
@@ -12,6 +12,10 @@ Description: 	Interface is used to configure RDMA-cable HCAs in respect to
 		for this HCA has to be created:
 		mkdir -p /config/rdma_cm/<hca>
 
+		Note: configfs has no network namespace support, so this
+		interface cannot represent two devices that share a name in
+		different network namespaces (possible in exclusive netns mode).
+
 
 What: 		/config/rdma_cm/<hca>/ports/<port-num>/default_roce_mode
 Date: 		November 29, 2015
drivers/infiniband/core/cma_configfs.c +4 -0
index 891e52afb8f4..c389d4e37b6b 100644
@@ -65,6 +65,10 @@ static struct cma_dev_port_group *to_dev_port_group(struct config_item *item)
 	return container_of(group, struct cma_dev_port_group, group);
 }
 
+/*
+ * configfs is not net namespace aware, so a name shared by devices in
+ * different namespaces resolves to the first match here.
+ */
 static bool filter_by_name(struct ib_device *ib_dev, void *cookie)
 {
 	return !strcmp(dev_name(&ib_dev->dev), cookie);
Signature · 1 line
2.54.0
author
#14 lore raw

PATCH v2 10/14 RDMA/core: Document the SELinux ibendport net namespace limitation

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Document that SELinux ibendport labels use a global (device name, port) key, so same-named RDMA devices in different net namespaces share a label.

---
 drivers/infiniband/core/security.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/infiniband/core/security.c +6 -0
index 9af31d1d9d70..a82c46965416 100644
@@ -700,6 +700,12 @@ int ib_mad_agent_security_setup(struct ib_mad_agent *agent,
 	if (qp_type != IB_QPT_SMI)
 		return 0;
 
+	/*
+	 * SELinux labels an endport by (device name, port) from a global
+	 * policy. If devices in different net namespaces share a name, they get
+	 * the same label; distinguishing them would need net namespace support
+	 * in the policy language and tooling.
+	 */
 	spin_lock(&mad_agent_list_lock);
 	ret = security_ib_endport_manage_subnet(agent->security,
 						dev_name(&agent->device->dev),
Signature · 1 line
2.54.0
author
#15 lore raw

PATCH v2 11/14 RDMA/core: Make device names unique per net namespace

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Use rdma_dev_access_netns() to scope RDMA device name lookup and "%d" name allocation to the relevant net namespace. Keep shared mode and CONFIG_NET_NS=n behaviour system-wide.

---
 drivers/infiniband/core/device.c | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)
drivers/infiniband/core/device.c +15 -7
index bb02640239d7..4c75078448a3 100644
@@ -129,7 +129,7 @@ static DECLARE_RWSEM(rdma_nets_rwsem);
 bool ib_devices_shared_netns = true;
 module_param_named(netns_mode, ib_devices_shared_netns, bool, 0444);
 MODULE_PARM_DESC(netns_mode,
-		 "Share device among net namespaces; default=1 (shared)");
+		 "Share device among net namespaces; default=1 (shared). In exclusive mode device names are unique per net namespace");
 /**
  * rdma_dev_access_netns() - Return whether an rdma device can be accessed
  *			     from a specified net namespace or not.
@@ -359,7 +359,8 @@ static struct ib_device *__ib_device_get_by_name(const char *name,
 	unsigned long index;
 
 	xa_for_each (&devices, index, device)
-		if (!strcmp(name, dev_name(&device->dev)))
+		if (rdma_dev_access_netns(device, net) &&
+		    !strcmp(name, dev_name(&device->dev)))
 			return device;
 
 	return NULL;
@@ -437,7 +438,11 @@ int ib_device_set_dim(struct ib_device *ibdev, u8 use_dim)
 	return 0;
 }
 
-/* Pick a free index for the '%d' style @name pattern. */
+/*
+ * Pick a free index for the '%d' style @name pattern within net namespace
+ * @net. Returns the index on success or a negative errno. The caller builds
+ * the final unique device name from the returned index.
+ */
 static int __alloc_name_id(struct net *net, const char *name,
 			   const struct ib_device *skip)
 {
@@ -452,7 +457,7 @@ static int __alloc_name_id(struct net *net, const char *name,
 	xa_for_each (&devices, index, device) {
 		char buf[IB_DEVICE_NAME_MAX];
 
-		if (device == skip)
+		if (device == skip || !rdma_dev_access_netns(device, net))
 			continue;
 		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 			continue;
@@ -1233,7 +1238,8 @@ static __net_init int rdma_dev_init_net(struct net *net)
 }
 
 /*
- * Assign the unique string device name and the unique device index. This is
+ * Assign the unique string device name and the unique device index. The device
+ * name is unique within the net namespace the device is assigned to. This is
  * undone by ib_dealloc_device.
  */
 static int assign_name(struct ib_device *device, const char *name)
@@ -1416,8 +1422,9 @@ static void ib_device_notify_register(struct ib_device *device)
 /**
  * ib_register_device - Register an IB device with IB core
  * @device: Device to register
- * @name: unique string device name. This may include a '%' which will
- * 	  cause a unique index to be added to the passed device name.
+ * @name: device name, unique within the device's net namespace. This may
+ *	  include a '%' which will cause a unique index to be added to the
+ *	  passed device name.
  * @dma_device: pointer to a DMA-capable device. If %NULL, then the IB
  *	        device will be used. In this case the caller should fully
  *		setup the ibdev for DMA. This usually means using dma_virt_ops.
@@ -1708,6 +1715,7 @@ static bool rdma_dev_name_in_netns(struct ib_device *skip, struct net *net,
 
 	xa_for_each(&devices, index, device)
 		if (device != skip &&
+		    rdma_dev_access_netns(device, net) &&
 		    !strcmp(name, dev_name(&device->dev)))
 			return true;
 
Signature · 1 line
2.54.0
#16 lore raw

Re: [PATCH rdma-next v2 11/14] RDMA/core: Make device names unique per net namespace

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Format String Vulnerability / Arbitrary Memory Write --

--- Patch [11]: [PATCH rdma-next v2 11/14] RDMA/core: Make device names unique per net namespace --- commit db3597c8065e66f96858593e01ba04bca7a7d075
Author: Jiri Pirko <[email protected]>

RDMA/core: Make device names unique per net namespace

This commit updates RDMA device name lookups and index allocations to be scoped within the relevant net namespace. It uses rdma_dev_access_netns() to ensure name uniqueness per namespace while maintaining system-wide behavior in shared mode.

Quoted 3 lines
drivers/infiniband/core/device.c +0 -0

[ ... ]

drivers/infiniband/core/device.c static int __alloc_name_id(struct net *net, const char *name, · line 463

Quoted 9 lines
drivers/infiniband/core/device.c +1 -1
@@ -452,7 +457,7 @@ static int __alloc_name_id(struct net *net, const char *name,
 	xa_for_each (&devices, index, device) {
 		char buf[IB_DEVICE_NAME_MAX];
 
-		if (device == skip)
+		if (device == skip || !rdma_dev_access_netns(device, net))
 			continue;
 		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 			continue;

[Severity: Critical]
This isn't a bug introduced by this patch, but is it possible that the sscanf call here is vulnerable to a format string injection?

Since name can be user-supplied via the RDMA_NLDEV_CMD_NEWDEV netlink command in nldev_newdev() and is copied without checking for '%' characters:

drivers/infiniband/core/nldev.c:nldev_newdev() {
        ...
        nla_strscpy(name, tb[RDMA_NLDEV_ATTR_DEV_NAME], sizeof(name));
        ...
}

The string is passed down through ib_register_device() and assign_name() into __alloc_name_id(). Could a user with CAP_NET_ADMIN supply a malicious format string containing format specifiers like %n?

If an attacker sets up a device with a matching prefix, sscanf would parse the attacker's format string and write to memory addresses beyond the supplied &i pointer.

Since CAP_NET_ADMIN can be acquired within a user namespace, could this allow unprivileged local attackers to corrupt memory and trigger an arbitrary memory write?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=11
author
#17 lore raw

Re: [PATCH rdma-next v2 11/14] RDMA/core: Make device names unique per net namespace

Quoted 27 lines · Wed, Jul 15, 2026 at 04:30:44PM +0200, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Format String Vulnerability / Arbitrary Memory Write --

--- Patch [11]: [PATCH rdma-next v2 11/14] RDMA/core: Make device names unique per net namespace --- commit db3597c8065e66f96858593e01ba04bca7a7d075
Author: Jiri Pirko <[email protected]>

RDMA/core: Make device names unique per net namespace

This commit updates RDMA device name lookups and index allocations to be scoped within the relevant net namespace. It uses rdma_dev_access_netns() to ensure name uniqueness per namespace while maintaining system-wide behavior in shared mode.

Quoted 3 lines
drivers/infiniband/core/device.c +0 -0

[ ... ]

Quoted 4 lines
drivers/infiniband/core/device.c +0 -0
@@ -452,7 +457,7 @@ static int __alloc_name_id(struct net *net, const char *name,
 	xa_for_each (&devices, index, device) {
 		char buf[IB_DEVICE_NAME_MAX];
 
Quoted 9 lines

drivers/infiniband/core/device.c static int __alloc_name_id(struct net *net, const char *name, · line 463

Quoted 5 lines
drivers/infiniband/core/device.c +1 -1
-		if (device == skip)
+		if (device == skip || !rdma_dev_access_netns(device, net))
 			continue;
 		if (sscanf(dev_name(&device->dev), name, &i) != 1)
 			continue;

[Severity: Critical]
This isn't a bug introduced by this patch, but is it possible that the sscanf call here is vulnerable to a format string injection?

Unrelated to this patch/set. Ignoring.

author
#18 lore raw

PATCH v2 12/14 RDMA/rxe: Allow queue VMAs to outlive ucontexts

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Prepare queue mappings for asynchronous ucontext disassociation during device disable. Rely on the VMA page references to preserve mapped memory until the final unmap.

---
v1->v2:
- new patch

---
 drivers/infiniband/sw/rxe/rxe_mmap.c | 35 ++--------------------------
 1 file changed, 2 insertions(+), 33 deletions(-)
drivers/infiniband/sw/rxe/rxe_mmap.c +2 -33
index 7f723a2f3700..a4ead89ccbd3 100644
@@ -11,7 +11,6 @@
 
 #include "rxe.h"
 #include "rxe_loc.h"
-#include "rxe_queue.h"
 
 void rxe_mmap_release(struct kref *ref)
 {
@@ -30,29 +29,6 @@ void rxe_mmap_release(struct kref *ref)
 	kfree(ip);
 }
 
-/*
- * open and close keep track of how many times the memory region is mapped,
- * to avoid releasing it.
- */
-static void rxe_vma_open(struct vm_area_struct *vma)
-{
-	struct rxe_mmap_info *ip = vma->vm_private_data;
-
-	kref_get(&ip->ref);
-}
-
-static void rxe_vma_close(struct vm_area_struct *vma)
-{
-	struct rxe_mmap_info *ip = vma->vm_private_data;
-
-	kref_put(&ip->ref, rxe_mmap_release);
-}
-
-static const struct vm_operations_struct rxe_vm_ops = {
-	.open = rxe_vma_open,
-	.close = rxe_vma_close,
-};
-
 /**
  * rxe_mmap - create a new mmap region
  * @context: the IB user context of the process making the mmap() call
@@ -106,17 +82,10 @@ int rxe_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 	list_del_init(&ip->pending_mmaps);
 	spin_unlock_bh(&rxe->pending_lock);
 
-	vma->vm_ops = &rxe_vm_ops;
-	vma->vm_private_data = ip;
-
 	ret = remap_vmalloc_range(vma, ip->obj, 0);
-	if (ret) {
-		vma->vm_private_data = NULL;
-		vma->vm_ops = NULL;
-		kref_put(&ip->ref, rxe_mmap_release);
+	kref_put(&ip->ref, rxe_mmap_release);
+	if (ret)
 		rxe_dbg_dev(rxe, "err %d from remap_vmalloc_range\n", ret);
-		goto done;
-	}
 
 done:
 	return ret;
Signature · 1 line
2.54.0
author
#19 lore raw

PATCH v2 13/14 RDMA/rxe: Implement disassociate_ucontext callback

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Implement an empty disassociate_ucontext() callback so the RDMA core can move rxe devices between net namespaces. The core requires this callback to reset user contexts without waiting for userspace.

rxe needs no teardown here: its user-mapped queues live in reference-counted vmalloc memory (see rxe_mmap.c) that stays valid while userspace holds the mappings.

---
 drivers/infiniband/sw/rxe/rxe_verbs.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/infiniband/sw/rxe/rxe_verbs.c +5 -0
index 1ec130fee8ea..6eb10d2f0653 100644
@@ -240,6 +240,10 @@ static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
 		rxe_err_uc(uc, "cleanup failed, err = %d\n", err);
 }
 
+static void rxe_disassociate_ucontext(struct ib_ucontext *ibuc)
+{
+}
+
 /* pd */
 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 {
@@ -1478,6 +1482,7 @@ static const struct ib_device_ops rxe_dev_ops = {
 	.destroy_srq = rxe_destroy_srq,
 	.detach_mcast = rxe_detach_mcast,
 	.device_group = &rxe_attr_group,
+	.disassociate_ucontext = rxe_disassociate_ucontext,
 	.enable_driver = rxe_enable_driver,
 	.get_dma_mr = rxe_get_dma_mr,
 	.get_hw_stats = rxe_ib_get_hw_stats,
Signature · 1 line
2.54.0
author
#20 lore raw

PATCH v2 14/14 RDMA/selftests: Add rxe_netns_names test

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Add a kselftest script that exercises per-netns RDMA device naming with RXE. Cover duplicate names across namespaces, move conflict handling, move-with-rename, and same-namespace rename requests.

---
v1->v2:
- fixed ktap_set_plan
- s/RXE_A/RXE_SAME/ in dup rename

---
 tools/testing/selftests/rdma/Makefile         |   3 +-
 tools/testing/selftests/rdma/config           |   2 +
 .../testing/selftests/rdma/rxe_netns_names.sh | 282 ++++++++++++++++++
 3 files changed, 286 insertions(+), 1 deletion(-)
 create mode 100755 tools/testing/selftests/rdma/rxe_netns_names.sh
tools/testing/selftests/rdma/Makefile +2 -1
index 07af7f15c1bf..a91c14c45006 100644
@@ -3,6 +3,7 @@ TEST_PROGS := rxe_rping_between_netns.sh \
 		rxe_ipv6.sh \
 		rxe_socket_with_netns.sh \
 		rxe_test_NETDEV_UNREGISTER.sh \
-		rxe_sent_rcvd_bytes.sh
+		rxe_sent_rcvd_bytes.sh \
+		rxe_netns_names.sh
 
 include ../lib.mk
tools/testing/selftests/rdma/config +2 -0
index 4ffb814e253b..e1ff54ec0f57 100644
@@ -1,3 +1,5 @@
 CONFIG_TUN
 CONFIG_VETH
+CONFIG_DUMMY
+CONFIG_NET_NS
 CONFIG_RDMA_RXE
tools/testing/selftests/rdma/rxe_netns_names.sh new file +282 -0

282 lines not on this page · raw message

Signature · 1 line
2.54.0
#21 lore raw

Re: [PATCH rdma-next v2 14/14] RDMA/selftests: Add rxe_netns_names test

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [Medium] Asynchronous RXE device deletion causes race conditions and test failures. - [Medium]…

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Medium] Asynchronous RXE device deletion causes race conditions and test failures.
- [Medium] Asynchronous network namespace deletion causes test teardown to permanently leak the exclusive netns mode. --

--- Patch [14]: [PATCH rdma-next v2 14/14] RDMA/selftests: Add rxe_netns_names test --- commit dcc30384e696a4047f319e3b7139b30b3ddc09f4
Author: Jiri Pirko <[email protected]>

RDMA/selftests: Add rxe_netns_names test

Add a kselftest script that exercises per-netns RDMA device naming with RXE. Cover duplicate names across namespaces, move conflict handling, move-with-rename, and same-namespace rename requests.

Quoted 4 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +0 -0
new file mode 100755

[ ... ]

tools/testing/selftests/rdma/rxe_netns_names.sh line 129

Quoted 5 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +5 -0
+cleanup_devs()
+{
+	ip link del "$DUMMY_A" 2>/dev/null
+	ip link del "$DUMMY_B" 2>/dev/null
+}

[Severity: Medium]
Will this cause race conditions between test iterations?

The ip link del command returns as soon as the netdev reference is dropped, but the rdma_rxe driver handles the NETDEV_UNREGISTER event by scheduling an asynchronous unregistration of the RDMA device via ib_unregister_device_queued().

Could the next test iteration fail with -EEXIST when attempting to create or rename a device to the exact same name (like $RXE_SAME) if the asynchronous device_del() from the previous test hasn't finished?

[ ... ]

tools/testing/selftests/rdma/rxe_netns_names.sh line 154

Quoted 10 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +10 -0
+cleanup()
+{
+	cleanup_devs
+
+	ip netns del "$NS1" 2>/dev/null
+	ip netns del "$NS2" 2>/dev/null
+
+	if [ "$MODE_CHANGED" -eq 1 ]; then
+		rdma system set netns "$OLD_MODE" 2>/dev/null
+	fi

[Severity: Medium]
Does this cleanup sequence permanently leak the exclusive netns mode?

The ip netns del command initiates asynchronous teardown by queuing net_cleanup_work. If rdma system set netns "$OLD_MODE" is executed immediately afterwards while the namespaces are still in the process of being destroyed, the kernel might reject the mode change with -EBUSY.

Since the script redirects errors to /dev/null, it seems the system's RDMA subsystem could permanently be left in exclusive mode instead of being restored to its original shared mode.

Quoted 5 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +5 -0
+
+	for m in "${MODS[@]}"; do
+		modprobe -r "$m" 2>/dev/null
+	done
+}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=14
author
#22 lore raw

Re: [PATCH rdma-next v2 14/14] RDMA/selftests: Add rxe_netns_names test

Quoted 31 lines · Wed, Jul 15, 2026 at 04:30:45PM +0200, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Medium] Asynchronous RXE device deletion causes race conditions and test failures.
- [Medium] Asynchronous network namespace deletion causes test teardown to permanently leak the exclusive netns mode. --

--- Patch [14]: [PATCH rdma-next v2 14/14] RDMA/selftests: Add rxe_netns_names test --- commit dcc30384e696a4047f319e3b7139b30b3ddc09f4
Author: Jiri Pirko <[email protected]>

RDMA/selftests: Add rxe_netns_names test

Add a kselftest script that exercises per-netns RDMA device naming with RXE. Cover duplicate names across namespaces, move conflict handling, move-with-rename, and same-namespace rename requests.

Quoted 4 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +0 -0
new file mode 100755

[ ... ]

tools/testing/selftests/rdma/rxe_netns_names.sh line 129

Quoted 5 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +5 -0
+cleanup_devs()
+{
+	ip link del "$DUMMY_A" 2>/dev/null
+	ip link del "$DUMMY_B" 2>/dev/null
+}

[Severity: Medium]
Will this cause race conditions between test iterations?

Quoted 8 lines

The ip link del command returns as soon as the netdev reference is dropped, but the rdma_rxe driver handles the NETDEV_UNREGISTER event by scheduling an asynchronous unregistration of the RDMA device via ib_unregister_device_queued().

Could the next test iteration fail with -EEXIST when attempting to create or rename a device to the exact same name (like $RXE_SAME) if the asynchronous device_del() from the previous test hasn't finished?

will be fixed in v3

Quoted 17 lines

[ ... ]

tools/testing/selftests/rdma/rxe_netns_names.sh line 154

Quoted 10 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +10 -0
+cleanup()
+{
+	cleanup_devs
+
+	ip netns del "$NS1" 2>/dev/null
+	ip netns del "$NS2" 2>/dev/null
+
+	if [ "$MODE_CHANGED" -eq 1 ]; then
+		rdma system set netns "$OLD_MODE" 2>/dev/null
+	fi

[Severity: Medium]
Does this cleanup sequence permanently leak the exclusive netns mode?

Quoted 8 lines

The ip netns del command initiates asynchronous teardown by queuing net_cleanup_work. If rdma system set netns "$OLD_MODE" is executed immediately afterwards while the namespaces are still in the process of being destroyed, the kernel might reject the mode change with -EBUSY.

Since the script redirects errors to /dev/null, it seems the system's RDMA subsystem could permanently be left in exclusive mode instead of being restored to its original shared mode.

will be fixed in v3

Quoted 9 lines
Quoted 6 lines
tools/testing/selftests/rdma/rxe_netns_names.sh new file +5 -0
+
+	for m in "${MODS[@]}"; do
+		modprobe -r "$m" 2>/dev/null
+	done
+}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=14