From: Gui-Dong Han
> Sent: 22 December 2023 10:55
>
> In {conn,adv}_min_interval_set():
> if (val < ... || val > ... || val > hdev->le_{conn,adv}_max_interval)
> return -EINVAL;
> hci_dev_lock(hdev);
> hdev->le_{conn,adv}_min_interval = val;
> hci_dev_unlock(hdev);
>
> In {conn,adv}_max_interval_set():
> if (val < ... || val > ... || val < hdev->le_{conn,adv}_min_interval)
> return -EINVAL;
> hci_dev_lock(hdev);
> hdev->le_{conn,adv}_max_interval
> hci_dev_unlock(hdev);
>
> The atomicity violation occurs due to concurrent execution of set_min and
> set_max funcs which may lead to inconsistent reads and writes of the min
> value and the max value. The checks for value validity are ineffective as
> the min/max values could change immediately after being checked, raising
> the risk of the min value being greater than the max value and causing
> invalid settings.
>
> This possible bug is found by an experimental static analysis tool
> developed by our team, BassCheck[1]. This tool analyzes the locking APIs
> to extract function pairs that can be concurrently executed, and then
> analyzes the instructions in the paired functions to identify possible
> concurrency bugs including data races and atomicity violations. The above
> possible bug is reported when our tool analyzes the source code of
> Linux 5.17.
Your static analysis tool is basically broken.
The only possible issues are if the accesses aren't atomic.
In practise they always will be but using READ_ONCE() and
WRITE_ONCE() would make that certain.
The lock sequence:
> hci_dev_lock(hdev);
> hdev->le_conn_min_interval = val;
> hci_dev_unlock(hdev);
is pretty pointless - is doesn't 'lock' two+ things together.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
Hi,
Thanks for your feedback. Let me clarify the potential issue with
concurrent execution of setmax and setmin functions. Consider a scenario
where setmin writes a new, valid 'min' value, and concurrently, setmax
writes a value that is greater than the old 'min' but smaller than the
new 'min'. In this case, setmax might check against the old 'min' value
(before acquiring the lock) but write its value after the 'min' has been
updated by setmin. This leads to a situation where the 'max' value ends
up being smaller than the 'min' value, which is an inconsistency.
Regarding the lock sequence you mentioned, it's indeed from the original
code. My patch aims to include the validity checks within the
lock/unlock sequence to prevent the described race condition.
Thanks,
Han
On 22/12/2023 下午7:41, David Laight wrote:
> From: Gui-Dong Han
>> Sent: 22 December 2023 10:55
>>
>> In {conn,adv}_min_interval_set():
>> if (val < ... || val > ... || val > hdev->le_{conn,adv}_max_interval)
>> return -EINVAL;
>> hci_dev_lock(hdev);
>> hdev->le_{conn,adv}_min_interval = val;
>> hci_dev_unlock(hdev);
>>
>> In {conn,adv}_max_interval_set():
>> if (val < ... || val > ... || val < hdev->le_{conn,adv}_min_interval)
>> return -EINVAL;
>> hci_dev_lock(hdev);
>> hdev->le_{conn,adv}_max_interval
>> hci_dev_unlock(hdev);
>>
>> The atomicity violation occurs due to concurrent execution of set_min and
>> set_max funcs which may lead to inconsistent reads and writes of the min
>> value and the max value. The checks for value validity are ineffective as
>> the min/max values could change immediately after being checked, raising
>> the risk of the min value being greater than the max value and causing
>> invalid settings.
>>
>> This possible bug is found by an experimental static analysis tool
>> developed by our team, BassCheck[1]. This tool analyzes the locking APIs
>> to extract function pairs that can be concurrently executed, and then
>> analyzes the instructions in the paired functions to identify possible
>> concurrency bugs including data races and atomicity violations. The above
>> possible bug is reported when our tool analyzes the source code of
>> Linux 5.17.
> Your static analysis tool is basically broken.
>
> The only possible issues are if the accesses aren't atomic.
> In practise they always will be but using READ_ONCE() and
> WRITE_ONCE() would make that certain.
>
> The lock sequence:
>> hci_dev_lock(hdev);
>> hdev->le_conn_min_interval = val;
>> hci_dev_unlock(hdev);
> is pretty pointless - is doesn't 'lock' two+ things together.
>
> David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
>