The patch below does not apply to the 6.12-stable tree. If someone wants it applied there, or to any other stable or longterm tree, then please email the backport, including the original git commit id to stable@vger.kernel.org.
To reproduce the conflict and resubmit, you may use the following commands:
git fetch https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ linux-6.12.y git checkout FETCH_HEAD git cherry-pick -x 59b33fab4ca4d7dacc03367082777627e05d0323 # <resolve conflicts, build, test, etc.> git commit -s git send-email --to 'stable@vger.kernel.org' --in-reply-to '2025081230-dentist-cautious-7bb4@gregkh' --subject-prefix 'PATCH 6.12.y' HEAD^..
Possible dependencies:
thanks,
greg k-h
------------------ original commit in Linus's tree ------------------
From 59b33fab4ca4d7dacc03367082777627e05d0323 Mon Sep 17 00:00:00 2001 From: Wang Zhaolong wangzhaolong@huaweicloud.com Date: Thu, 17 Jul 2025 21:29:26 +0800 Subject: [PATCH] smb: client: fix netns refcount leak after net_passive changes
After commit 5c70eb5c593d ("net: better track kernel sockets lifetime"), kernel sockets now use net_passive reference counting. However, commit 95d2b9f693ff ("Revert "smb: client: fix TCP timers deadlock after rmmod"") restored the manual socket refcount manipulation without adapting to this new mechanism, causing a memory leak.
The issue can be reproduced by[1]: 1. Creating a network namespace 2. Mounting and Unmounting CIFS within the namespace 3. Deleting the namespace
Some memory leaks may appear after a period of time following step 3.
unreferenced object 0xffff9951419f6b00 (size 256): comm "ip", pid 447, jiffies 4294692389 (age 14.730s) hex dump (first 32 bytes): 1b 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ 00 00 00 00 00 00 00 00 80 77 c2 44 51 99 ff ff .........w.DQ... backtrace: __kmem_cache_alloc_node+0x30e/0x3d0 __kmalloc+0x52/0x120 net_alloc_generic+0x1d/0x30 copy_net_ns+0x86/0x200 create_new_namespaces+0x117/0x300 unshare_nsproxy_namespaces+0x60/0xa0 ksys_unshare+0x148/0x360 __x64_sys_unshare+0x12/0x20 do_syscall_64+0x59/0x110 entry_SYSCALL_64_after_hwframe+0x78/0xe2 ... unreferenced object 0xffff9951442e7500 (size 32): comm "mount.cifs", pid 475, jiffies 4294693782 (age 13.343s) hex dump (first 32 bytes): 40 c5 38 46 51 99 ff ff 18 01 96 42 51 99 ff ff @.8FQ......BQ... 01 00 00 00 6f 00 c5 07 6f 00 d8 07 00 00 00 00 ....o...o....... backtrace: __kmem_cache_alloc_node+0x30e/0x3d0 kmalloc_trace+0x2a/0x90 ref_tracker_alloc+0x8e/0x1d0 sk_alloc+0x18c/0x1c0 inet_create+0xf1/0x370 __sock_create+0xd7/0x1e0 generic_ip_connect+0x1d4/0x5a0 [cifs] cifs_get_tcp_session+0x5d0/0x8a0 [cifs] cifs_mount_get_session+0x47/0x1b0 [cifs] dfs_mount_share+0xfa/0xa10 [cifs] cifs_mount+0x68/0x2b0 [cifs] cifs_smb3_do_mount+0x10b/0x760 [cifs] smb3_get_tree+0x112/0x2e0 [cifs] vfs_get_tree+0x29/0xf0 path_mount+0x2d4/0xa00 __se_sys_mount+0x165/0x1d0
Root cause: When creating kernel sockets, sk_alloc() calls net_passive_inc() for sockets with sk_net_refcnt=0. The CIFS code manually converts kernel sockets to user sockets by setting sk_net_refcnt=1, but doesn't call the corresponding net_passive_dec(). This creates an imbalance in the net_passive counter, which prevents the network namespace from being destroyed when its last user reference is dropped. As a result, the entire namespace and all its associated resources remain allocated.
Timeline of patches leading to this issue: - commit ef7134c7fc48 ("smb: client: Fix use-after-free of network namespace.") in v6.12 fixed the original netns UAF by manually managing socket refcounts - commit e9f2517a3e18 ("smb: client: fix TCP timers deadlock after rmmod") in v6.13 attempted to use kernel sockets but introduced TCP timer issues - commit 5c70eb5c593d ("net: better track kernel sockets lifetime") in v6.14-rc5 introduced the net_passive mechanism with sk_net_refcnt_upgrade() for proper socket conversion - commit 95d2b9f693ff ("Revert "smb: client: fix TCP timers deadlock after rmmod"") in v6.15-rc3 reverted to manual refcount management without adapting to the new net_passive changes
Fix this by using sk_net_refcnt_upgrade() which properly handles the net_passive counter when converting kernel sockets to user sockets.
Link: https://bugzilla.kernel.org/show_bug.cgi?id=220343 [1] Fixes: 95d2b9f693ff ("Revert "smb: client: fix TCP timers deadlock after rmmod"") Cc: stable@vger.kernel.org Reviewed-by: Kuniyuki Iwashima kuniyu@google.com Reviewed-by: Enzo Matsumiya ematsumiya@suse.de Signed-off-by: Wang Zhaolong wangzhaolong@huaweicloud.com Signed-off-by: Steve French stfrench@microsoft.com
diff --git a/fs/smb/client/connect.c b/fs/smb/client/connect.c index 205f547ca49e..5eec8957f2a9 100644 --- a/fs/smb/client/connect.c +++ b/fs/smb/client/connect.c @@ -3362,18 +3362,15 @@ generic_ip_connect(struct TCP_Server_Info *server) struct net *net = cifs_net_ns(server); struct sock *sk;
- rc = __sock_create(net, sfamily, SOCK_STREAM, - IPPROTO_TCP, &server->ssocket, 1); + rc = sock_create_kern(net, sfamily, SOCK_STREAM, + IPPROTO_TCP, &server->ssocket); if (rc < 0) { cifs_server_dbg(VFS, "Error %d creating socket\n", rc); return rc; }
sk = server->ssocket->sk; - __netns_tracker_free(net, &sk->ns_tracker, false); - sk->sk_net_refcnt = 1; - get_net_track(net, &sk->ns_tracker, GFP_KERNEL); - sock_inuse_add(net, 1); + sk_net_refcnt_upgrade(sk);
/* BB other socket options to set KEEPALIVE, NODELAY? */ cifs_dbg(FYI, "Socket created\n");