Message-ID: <CAHS8izMrN4+UuoRy3zUS0-2KJGfUhRVxyeJHEn81VX=9TdjKcg@mail.gmail.com>
Date: Tue, 15 Apr 2025 11:22:44 -0700
From: Mina Almasry <almasrymina@...gle.com>
To: Taehee Yoo <ap420073@...il.com>
Cc: davem@...emloft.net, kuba@...nel.org, pabeni@...hat.com, 
	edumazet@...gle.com, andrew+netdev@...n.ch, horms@...nel.org, 
	asml.silence@...il.com, dw@...idwei.uk, sdf@...ichev.me, skhawaja@...gle.com, 
	simona.vetter@...ll.ch, kaiyuanz@...gle.com, netdev@...r.kernel.org
Subject: Re: [PATCH net] net: devmem: fix kernel panic when socket close after
 module unload

On Tue, Apr 15, 2025 at 2:24 AM Taehee Yoo <ap420073@...il.com> wrote:
>
> Kernel panic occurs when a devmem TCP socket is closed after NIC module
> is unloaded.
>
> This is Devmem TCP unregistration scenarios. number is an order.
> (a)socket close    (b)pp destroy    (c)uninstall    result
> 1                  2                3               OK
> 1                  3                2               (d)Impossible
> 2                  1                3               OK
> 3                  1                2               (e)Kernel panic
> 2                  3                1               (d)Impossible
> 3                  2                1               (d)Impossible
>
> (a) netdev_nl_sock_priv_destroy() is called when devmem TCP socket is
>     closed.
> (b) page_pool_destroy() is called when the interface is down.
> (c) mp_ops->uninstall() is called when an interface is unregistered.
> (d) There is no scenario in mp_ops->uninstall() is called before
>     page_pool_destroy().
>     Because unregister_netdevice_many_notify() closes interfaces first
>     and then calls mp_ops->uninstall().
> (e) netdev_nl_sock_priv_destroy() accesses struct net_device.
>     But if the interface module has already been removed, net_device
>     pointer is invalid, so it causes kernel panic.
>
> In summary, there are only 3 possible scenarios.
>  A. sk close -> pp destroy -> uninstall.
>  B. pp destroy -> sk close -> uninstall.
>  C. pp destroy -> uninstall -> sk close.
>
> Case C is a kernel panic scenario.
>
> In order to fix this problem, it makes netdev_nl_sock_priv_destroy() do
> nothing if a module is already removed.
> The mp_ops->uninstall() handles these instead.
>
> The netdev_nl_sock_priv_destroy() iterates binding->list and releases
> them all with net_devmem_unbind_dmabuf().
> The net_devmem_unbind_dmabuf() has the below steps.
> 1. Delete binding from a list.
> 2. Call _net_mp_close_rxq() for all rxq's bound to a binding.
> 3. Call net_devmem_dmabuf_binding_put() to release resources.
>
> The mp_ops->uninstall() doesn't need to call _net_mp_close_rxq() because
> resources are already released properly when an interface is being down.
>
> From now on netdev_nl_sock_priv_destroy() will do nothing if a module
> has been removed because all bindings are removed from a list by
> mp_ops->uninstall().
>
> netdev_nl_sock_priv_destroy() internally sets mp_ops to NULL.
> So mp_ops->uninstall has not been called if devmem TCP socket was
> already closed.
>
> Tests:
> Scenario A:
>     ./ncdevmem -s 192.168.1.4 -c 192.168.1.2 -f $interface -l -p 8000 \
>         -v 7 -t 1 -q 1 &
>     pid=$!
>     sleep 10
>     ip link set $interface down
>     kill $pid
>     modprobe -rv $module
>
> Scenario B:
>     ./ncdevmem -s 192.168.1.4 -c 192.168.1.2 -f $interface -l -p 8000 \
>         -v 7 -t 1 -q 1 &
>     pid=$!
>     sleep 10
>     ip link set $interface down
>     kill $pid
>     modprobe -rv $module
>

Scenario A & B are exactly the same steps?

> Scenario C:
>     ./ncdevmem -s 192.168.1.4 -c 192.168.1.2 -f $interface -l -p 8000 \
>         -v 7 -t 1 -q 1 &
>     pid=$!
>     sleep 10
>     modprobe -rv $module
>     sleep 5
>     kill $pid
>
> Splat looks like:
> Oops: general protection fault, probably for non-canonical address 0xdffffc001fffa9f7: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN NOPTI
> KASAN: probably user-memory-access in range [0x00000000fffd4fb8-0x00000000fffd4fbf]
> CPU: 0 UID: 0 PID: 2041 Comm: ncdevmem Tainted: G    B   W           6.15.0-rc1+ #2 PREEMPT(undef)  0947ec89efa0fd68838b78e36aa1617e97ff5d7f
> Tainted: [B]=BAD_PAGE, [W]=WARN
> RIP: 0010:__mutex_lock (./include/linux/sched.h:2244 kernel/locking/mutex.c:400 kernel/locking/mutex.c:443 kernel/locking/mutex.c:605 kernel/locking/mutex.c:746)
> Code: ea 03 80 3c 02 00 0f 85 4f 13 00 00 49 8b 1e 48 83 e3 f8 74 6a 48 b8 00 00 00 00 00 fc ff df 48 8d 7b 34 48 89 fa 48 c1 ea 03 <0f> b6 f
> RSP: 0018:ffff88826f7ef730 EFLAGS: 00010203
> RAX: dffffc0000000000 RBX: 00000000fffd4f88 RCX: ffffffffaa9bc811
> RDX: 000000001fffa9f7 RSI: 0000000000000008 RDI: 00000000fffd4fbc
> RBP: ffff88826f7ef8b0 R08: 0000000000000000 R09: ffffed103e6aa1a4
> R10: 0000000000000007 R11: ffff88826f7ef442 R12: fffffbfff669f65e
> R13: ffff88812a830040 R14: ffff8881f3550d20 R15: 00000000fffd4f88
> FS:  0000000000000000(0000) GS:ffff888866c05000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 0000563bed0cb288 CR3: 00000001a7c98000 CR4: 00000000007506f0
> PKRU: 55555554
> Call Trace:
> <TASK>
>  ...
>  netdev_nl_sock_priv_destroy (net/core/netdev-genl.c:953 (discriminator 3))

Line 953 is:

netdev_lock(dev);

Which was introduced by:

commit 42f342387841 ("net: fix use-after-free in the
netdev_nl_sock_priv_destroy()") and rolling back a few fixes, it's
really introduced by commit 1d22d3060b9b ("net: drop rtnl_lock for
queue_mgmt operations").

My first question, does this issue still reproduce if you remove the
per netdev locking and go back to relying on rtnl_locking? Or do we
crash somewhere else in net_devmem_unbind_dmabuf? If so, where?
Looking through the rest of the unbinding code, it's not clear to me
any of it actually uses dev, so it may just be the locking...

>  genl_release (net/netlink/genetlink.c:653 net/netlink/genetlink.c:694 net/netlink/genetlink.c:705)
>  ...
>  netlink_release (net/netlink/af_netlink.c:737)
>  ...
>  __sock_release (net/socket.c:647)
>  sock_close (net/socket.c:1393)
>
> Fixes: 170aafe35cb9 ("netdev: support binding dma-buf to netdevice")
> Signed-off-by: Taehee Yoo <ap420073@...il.com>
> ---
>
> In order to test this patch, driver side implementation of devmem TCP[1]
> is needed to be applied.
>
> [1] https://lore.kernel.org/netdev/20250415052458.1260575-1-ap420073@gmail.com/T/#u
>
>  net/core/devmem.c | 5 +++++
>  1 file changed, 5 insertions(+)
>
> diff --git a/net/core/devmem.c b/net/core/devmem.c
> index 6e27a47d0493..8948796b0af5 100644
> --- a/net/core/devmem.c
> +++ b/net/core/devmem.c
> @@ -379,6 +379,11 @@ static void mp_dmabuf_devmem_uninstall(void *mp_priv,
>         xa_for_each(&binding->bound_rxqs, xa_idx, bound_rxq) {
>                 if (bound_rxq == rxq) {
>                         xa_erase(&binding->bound_rxqs, xa_idx);
> +
> +                       if (xa_empty(&binding->bound_rxqs)) {
> +                               list_del(&binding->list);
> +                               net_devmem_dmabuf_binding_put(binding);

On the surface, this fix looks completely unreviewable to be honest.
refcounting must be balanced. i.e. every put has a corresponding get,
otherwise there is a double free. I'm not sure which get you're
dropping here. I think this will cause a double put when the netlink
socket is closed?

--
Thanks,
Mina

Powered by blists - more mailing lists