Re: [PATCH -next,v2] serial: core: fix NULL/dangling port_dev on failed re-register

[cuigaosheng <cuigaosheng1@huawei.com>]

Sorry for the slow response, I was on holiday.

On 2026/9/24 22:40, Greg KH wrote:
> On Thu, Sep 24, 2026 at 09:11:47PM +0800, Gaosheng Cui wrote:
>> From: Cui GaoSheng <cuigaosheng1@huawei.com>
>>
>> serial8250_unregister_port() removes a port and then re-registers it
>> against the legacy serial8250 device when that device is still
>> present, ignoring the return value of uart_add_one_port(). When the
>> re-register fails, the error path of serial_core_register_port()
>> removes the port device but leaves port->port_dev dangling (or NULL
>> if the port device allocation itself failed). serial8250 matches
>> ports on ->dev, so such half-registered slots still reach
>> serial_core_unregister_port() on the next unbind:
> So this is abusing bind/unbind and not anything that normally happens in
> a system, right?

Yes, this was found by syzkaller, and the bind/unbind plus fail-nth fault injection
is just how it drives the error path, so it is not a normal workload.

>>    Oops: general protection fault ... SMP KASAN PTI
>>    KASAN: null-ptr-deref in range [0x40-0x47]
>>    RIP: serial_core_unregister_port+0xe5/0xa00
>>      serial8250_unregister_port+0x103
>>      serial8250_remove+0x61
>>      device_release_driver_internal+0x38c
>>      unbind_store+0xd1
>>      kernfs_fop_write_iter+0x329
>>      vfs_write+0x5d0
>>      ksys_write+0xf7
>>      do_syscall_64+0xdd
>>      entry_SYSCALL_64_after_hwframe+0x77
>>
>> Fix the error path to clear port->port_dev, and bail out of
>> serial_core_unregister_port() when port_dev is NULL. Read
>> port->port_dev under port_mutex, like all of its writers.
>>
>> Reject registering a port that is already registered: a duplicate
>> attempt would otherwise mark the live port dead and orphan its
>> serial core port device before failing in serial_core_add_one_port(),
>> and the port could then no longer be unregistered.
> How can registering a port that is already registered ever actually
> happen?

It can't, nothing in-tree registers the same port twice, the check was hardening
for a hypothetical driver bug from the v1 discussion, not for anything reachable.

>
>> No state slot restore is needed on the error path:
>> serial_core_add_one_port() has no failure exit after linking
>> state->uart_port, so a failed registration never links the state
>> slot in the first place.
>>
>> Fixes: 84a9582fd203 ("serial: core: Start managing serial controllers to enable runtime PM")
>> Assisted-by: GLM-5.3 OpenCode # fault analysis and fix rework
>> Signed-off-by: Cui GaoSheng <cuigaosheng1@huawei.com>
>> ---
>> Changes in v2:
>> - Drop the state slot restore from the registration error path.
>> - Move the port->port_dev read and the !port_dev bail in
>>    serial_core_unregister_port() under port_mutex.
>> - Reject registering a port that is already registered.
>>   drivers/tty/serial/serial_core.c | 37 +++++++++++++++++++++++++++++---
>>   1 file changed, 34 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
>> index 95774b0f1484..5651313a2468 100644
>> --- a/drivers/tty/serial/serial_core.c
>> +++ b/drivers/tty/serial/serial_core.c
>> @@ -3318,10 +3318,20 @@ static int serial_core_port_device_add(struct serial_ctrl_device *ctrl_dev,
>>   int serial_core_register_port(struct uart_driver *drv, struct uart_port *port)
>>   {
>>   	struct serial_ctrl_device *ctrl_dev, *new_ctrl_dev = NULL;
>> -	int ret;
>> +	int i, ret;
>>   
>>   	guard(mutex)(&port_mutex);
>>   
>> +	/*
>> +	 * Reject registration of a port that is already registered. Bail
>> +	 * out before taking any steps (UPF_DEAD, serial core devices) so
>> +	 * that a duplicate attempt cannot disturb the live registration.
>> +	 */
>> +	for (i = 0; i < drv->nr; i++) {
>> +		if (drv->state[i].uart_port == port)
>> +			return -EINVAL;
>> +	}
> Again, what normal code path attempts to register a port that is already
> registered?

It can't, nothing in-tree registers the same port twice, the check was hardening
for a hypothetical driver bug from the v1 discussion, not for anything reachable.

>> +
>>   	/*
>>   	 * Prevent serial_port_runtime_resume() from trying to use the port
>>   	 * until serial_core_add_one_port() has completed
>> @@ -3358,6 +3368,14 @@ int serial_core_register_port(struct uart_driver *drv, struct uart_port *port)
>>   
>>   err_unregister_port_dev:
>>   	serial_base_port_device_remove(port->port_dev);
>> +	/*
>> +	 * serial_base_port_device_remove() dropped the last reference to
>> +	 * the port device, so leave no dangling ->port_dev behind. No
>> +	 * state slot restore is needed: serial_core_add_one_port() has no
>> +	 * failure exit after linking state->uart_port, so a failed
>> +	 * registration never links the state slot.
>> +	 */
> Is this comment really needed?  LLMs love to do this, please don't.

It isn't, we can drop the comments.

>> +	port->port_dev = NULL;
>>   
>>   err_unregister_ctrl_dev:
>>   	serial_base_ctrl_device_remove(new_ctrl_dev);
>> @@ -3372,12 +3390,25 @@ int serial_core_register_port(struct uart_driver *drv, struct uart_port *port)
>>   void serial_core_unregister_port(struct uart_driver *drv, struct uart_port *port)
>>   {
>>   	struct device *phys_dev = port->dev;
>> -	struct serial_port_device *port_dev = port->port_dev;
>> -	struct serial_ctrl_device *ctrl_dev = serial_core_get_ctrl_dev(port_dev);
>> +	struct serial_port_device *port_dev;
>> +	struct serial_ctrl_device *ctrl_dev;
>>   	int ctrl_id = port->ctrl_id;
>>   
>>   	guard(mutex)(&port_mutex);
>>   
>> +	/*
>> +	 * A port may be unregistered even though it was never successfully
>> +	 * registered (e.g. an allocation failure aborted serial_core_register_port()
>> +	 * part-way).
> Ok, that might happen, but if it does, unregister isn't called, the port
> is just "dead".

It's true for normal drivers, but serial8250 removes ports by ->dev,
not by registration success, and re-registers them on every unbind
while ignoring the return value. A failed re-register therefore
still gets matched and removed by the next unbind - that is the
oops in the commit message. The leftover is a dangling pointer,
not a dead port; the fix is what makes it "just dead".

>> Drivers like serial8250 remove ports matching on ->dev, so
>> +	 * those half-registered ports still reach us with port_dev == NULL. Read
>> +	 * ->port_dev under port_mutex like all of its writers.
> How are you removing a port that failed to be added?
>
> I'm all for fixing real issues, but not fake ones.

This was found by syzkaller with fail-nth fault injection, so the
trigger is a fuzzer path, not a normal workload, but the oops is
from a real run.

I've sent a v3 with just the minimal fix, worth fixing, or too
theoretical to bother?

Thanks.

> thanks,
>
> greg k-h
>
> .