PDP contexts can be deleted through GTP_CMD_DELPDP or while the GTP network device is being unregistered. The latter is serialized by RTNL, but the generic-netlink delete path only holds RCU.
Running both paths concurrently can therefore make both paths delete the same PDP context. The issue was found through static analysis and reproduced on a KASAN-enabled kernel by a two-thread program racing GTP_CMD_DELPDP against RTM_DELLINK:
Oops: general protection fault, probably for non-canonical address KASAN: maybe wild-memory-access in range [0xdead000000000120-0xdead000000000127] RIP: gtp_genl_del_pdp+0x1c1/0x420 [gtp] RBP: dead000000000122
The second deletion dereferenced the poisoned hlist pprev pointer.
Serialize PDP context updates with a mutex shared by gtp_pdp_add(), gtp_genl_del_pdp(), and gtp_dellink().
Fixes: 459aa660eb1d ("gtp: add initial driver for datapath of GPRS Tunneling Protocol (GTP-U)") Assisted-by: Codex:gpt-5 Signed-off-by: Qing Ming a0yami@mailbox.org --- v3: - Rebase onto net/main.
v2: https://lore.kernel.org/netdev/20260808124322.5800-1-a0yami@mailbox.org/ - Use a dedicated mutex instead of RTNL. - Protect PDP add, delete, and link teardown with the same mutex.
drivers/net/gtp.c | 9 +++++++++ 1 file changed, 9 insertions(+)
diff --git a/drivers/net/gtp.c b/drivers/net/gtp.c index 9a12cc53da00..8ee516debeed 100644 --- a/drivers/net/gtp.c +++ b/drivers/net/gtp.c @@ -12,6 +12,7 @@ #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/module.h> +#include <linux/mutex.h> #include <linux/skbuff.h> #include <linux/udp.h> #include <linux/rculist.h> @@ -108,6 +109,7 @@ struct gtp_net { };
static u32 gtp_h_initval; +static DEFINE_MUTEX(gtp_pdp_lock);
static struct genl_family gtp_genl_family;
@@ -1555,9 +1557,11 @@ static void gtp_dellink(struct net_device *dev, struct list_head *head) struct pdp_ctx *pctx; int i;
+ mutex_lock(>p_pdp_lock); for (i = 0; i < gtp->hash_size; i++) hlist_for_each_entry_safe(pctx, next, >p->tid_hash[i], hlist_tid) pdp_context_delete(pctx); + mutex_unlock(>p_pdp_lock);
list_del(>p->list); unregister_netdevice_queue(dev, head); @@ -1833,6 +1837,8 @@ static struct pdp_ctx *gtp_pdp_add(struct gtp_dev *gtp, struct sock *sk, __be32 ms_addr; int family;
+ guard(mutex)(>p_pdp_lock); + version = nla_get_u32(info->attrs[GTPA_VERSION]);
family = nla_get_u8_default(info->attrs[GTPA_FAMILY], AF_INET); @@ -2134,6 +2140,8 @@ static int gtp_genl_del_pdp(struct sk_buff *skb, struct genl_info *info) if (!info->attrs[GTPA_VERSION]) return -EINVAL;
+ mutex_lock(>p_pdp_lock); + rcu_read_lock();
pctx = gtp_find_pdp(sock_net(skb->sk), info->attrs); @@ -2154,6 +2162,7 @@ static int gtp_genl_del_pdp(struct sk_buff *skb, struct genl_info *info)
out_unlock: rcu_read_unlock(); + mutex_unlock(>p_pdp_lock); return err; }
From: Qing Ming a0yami@mailbox.org Date: Tue, 11 Aug 2026 10:20:12 +0800
PDP contexts can be deleted through GTP_CMD_DELPDP or while the GTP network device is being unregistered. The latter is serialized by RTNL, but the generic-netlink delete path only holds RCU.
Running both paths concurrently can therefore make both paths delete the same PDP context. The issue was found through static analysis and reproduced on a KASAN-enabled kernel by a two-thread program racing GTP_CMD_DELPDP against RTM_DELLINK:
Oops: general protection fault, probably for non-canonical address KASAN: maybe wild-memory-access in range [0xdead000000000120-0xdead000000000127] RIP: gtp_genl_del_pdp+0x1c1/0x420 [gtp] RBP: dead000000000122
The second deletion dereferenced the poisoned hlist pprev pointer.
Serialize PDP context updates with a mutex shared by gtp_pdp_add(), gtp_genl_del_pdp(), and gtp_dellink().
Fixes: 459aa660eb1d ("gtp: add initial driver for datapath of GPRS Tunneling Protocol (GTP-U)") Assisted-by: Codex:gpt-5 Signed-off-by: Qing Ming a0yami@mailbox.org
v3:
- Rebase onto net/main.
v2: https://lore.kernel.org/netdev/20260808124322.5800-1-a0yami@mailbox.org/
- Use a dedicated mutex instead of RTNL.
- Protect PDP add, delete, and link teardown with the same mutex.
drivers/net/gtp.c | 9 +++++++++ 1 file changed, 9 insertions(+)
diff --git a/drivers/net/gtp.c b/drivers/net/gtp.c index 9a12cc53da00..8ee516debeed 100644 --- a/drivers/net/gtp.c +++ b/drivers/net/gtp.c @@ -12,6 +12,7 @@ #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/module.h> +#include <linux/mutex.h> #include <linux/skbuff.h> #include <linux/udp.h> #include <linux/rculist.h> @@ -108,6 +109,7 @@ struct gtp_net { };
static u32 gtp_h_initval; +static DEFINE_MUTEX(gtp_pdp_lock);
static struct genl_family gtp_genl_family;
@@ -1555,9 +1557,11 @@ static void gtp_dellink(struct net_device *dev, struct list_head *head) struct pdp_ctx *pctx; int i;
mutex_lock(>p_pdp_lock); for (i = 0; i < gtp->hash_size; i++) hlist_for_each_entry_safe(pctx, next, >p->tid_hash[i], hlist_tid) pdp_context_delete(pctx);
mutex_unlock(>p_pdp_lock);
list_del(>p->list); unregister_netdevice_queue(dev, head);
@@ -1833,6 +1837,8 @@ static struct pdp_ctx *gtp_pdp_add(struct gtp_dev *gtp, struct sock *sk, __be32 ms_addr; int family;
- guard(mutex)(>p_pdp_lock);
Please use the plain mutex_lock() / mutex_unlock().
---8<--- $ cat Documentation/process/maintainer-netdev.rst ... Using device-managed and cleanup.h constructs ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Netdev remains skeptical about promises of all "auto-cleanup" APIs, including even ``devm_`` helpers, historically. They are not the preferred style of implementation, merely an acceptable one.
Use of ``guard()`` is discouraged within any function longer than 20 lines, ``scoped_guard()`` is considered more readable. Using normal lock/unlock is still (weakly) preferred. ---8<---
On Tue, Aug 11, 2026 at 10:20:12AM +0800, Qing Ming wrote:
PDP contexts can be deleted through GTP_CMD_DELPDP or while the GTP network device is being unregistered. The latter is serialized by RTNL, but the generic-netlink delete path only holds RCU.
Running both paths concurrently can therefore make both paths delete the same PDP context. The issue was found through static analysis and reproduced on a KASAN-enabled kernel by a two-thread program racing GTP_CMD_DELPDP against RTM_DELLINK:
Oops: general protection fault, probably for non-canonical address KASAN: maybe wild-memory-access in range [0xdead000000000120-0xdead000000000127] RIP: gtp_genl_del_pdp+0x1c1/0x420 [gtp] RBP: dead000000000122
The second deletion dereferenced the poisoned hlist pprev pointer.
Serialize PDP context updates with a mutex shared by gtp_pdp_add(), gtp_genl_del_pdp(), and gtp_dellink().
Fixes: 459aa660eb1d ("gtp: add initial driver for datapath of GPRS Tunneling Protocol (GTP-U)") Assisted-by: Codex:gpt-5 Signed-off-by: Qing Ming a0yami@mailbox.org
v3:
- Rebase onto net/main.
v2: https://lore.kernel.org/netdev/20260808124322.5800-1-a0yami@mailbox.org/
- Use a dedicated mutex instead of RTNL.
- Protect PDP add, delete, and link teardown with the same mutex.
drivers/net/gtp.c | 9 +++++++++ 1 file changed, 9 insertions(+)
diff --git a/drivers/net/gtp.c b/drivers/net/gtp.c index 9a12cc53da00..8ee516debeed 100644 --- a/drivers/net/gtp.c +++ b/drivers/net/gtp.c
[...]
@@ -2134,6 +2140,8 @@ static int gtp_genl_del_pdp(struct sk_buff *skb, struct genl_info *info) if (!info->attrs[GTPA_VERSION]) return -EINVAL;
- mutex_lock(>p_pdp_lock);
- rcu_read_lock();
This rcu_read_lock() can go away after adding this new mutex.
pctx = gtp_find_pdp(sock_net(skb->sk), info->attrs); @@ -2154,6 +2162,7 @@ static int gtp_genl_del_pdp(struct sk_buff *skb, struct genl_info *info)
out_unlock: rcu_read_unlock();
This can go away too, for gtp_genl_del_pdp(), the mutex replaces this incorrect rcu read size lock.
Thanks.
- mutex_unlock(>p_pdp_lock); return err;
}
-- 2.53.0
在 2026/08/11 17:20 CST Pablo Neira Ayuso pablo@netfilter.org 写道: This can go away too, for gtp_genl_del_pdp(), the mutex replaces this incorrect rcu read size lock.
Thanks.
- mutex_unlock(>p_pdp_lock); return err;
}
-- 2.53.0
gtp_find_pdp() calls gtp_find_dev(), which uses dev_get_by_index_rcu(). Should I keep a narrow RCU read-side section around gtp_find_pdp(), while using the mutex to protect the PDP context?
Thanks, Qing
On Fri, Aug 14, 2026 at 12:31:59AM +0800, Ming Qing wrote:
在 2026/08/11 17:20 CST Pablo Neira Ayuso pablo@netfilter.org 写道: This can go away too, for gtp_genl_del_pdp(), the mutex replaces this incorrect rcu read size lock.
Thanks.
- mutex_unlock(>p_pdp_lock); return err;
}
-- 2.53.0
gtp_find_pdp() calls gtp_find_dev(), which uses dev_get_by_index_rcu(). Should I keep a narrow RCU read-side section around gtp_find_pdp(), while using the mutex to protect the PDP context?
Indeed, I forgot about gtp_dev uses rcu. Thanks for explaining.
osmocom-net-gprs@lists.osmocom.org
-
Kuniyuki Iwashima -
Ming Qing -
Pablo Neira Ayuso -
Qing Ming