Hi Anna and Trond,
I came across the following while reviewing atomic_inc_return() usages
that cast return value to unsigned
rpc_xprt_debugfs_register()'s atomic_inc_return() usage looks a bit odd.
- cur_id isn't initialized
- id = (unsigned int)atomic_inc_return(&cur_id);
Please note that id is int. Is it expected that cur_id could overflow?
Is there a maximum limit for this value?
thanks,
-- Shuah
On Wed, 2021-01-20 at 16:52 -0700, Shuah Khan wrote:
> Hi Anna and Trond,
>
> I came across the following while reviewing atomic_inc_return()
> usages
> that cast return value to unsigned
>
> rpc_xprt_debugfs_register()'s atomic_inc_return() usage looks a bit
> odd.
>
> - cur_id isn't initialized
> - id = (unsigned int)atomic_inc_return(&cur_id);
>
> Please note that id is int. Is it expected that cur_id could
> overflow?
> Is there a maximum limit for this value?
>
Yes, we do expect cur_id to eventually overflow (once you have created
2 billion RPC client instances), however the atomic increment
operations are expected to handle this correctly according to the
maintainers (I already asked them in a different context). Furthermore,
the code itself doesn't care about strict sequentiality. All it wants
from the counter is uniqueness, with that uniqueness condition actually
being enforced by the subsequent debugfs_create_file() call.
IOW: I don't think this is a real problem.
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
On 1/21/21 10:56 AM, Trond Myklebust wrote:
> On Wed, 2021-01-20 at 16:52 -0700, Shuah Khan wrote:
>> Hi Anna and Trond,
>>
>> I came across the following while reviewing atomic_inc_return()
>> usages
>> that cast return value to unsigned
>>
>> rpc_xprt_debugfs_register()'s atomic_inc_return() usage looks a bit
>> odd.
>>
>> - cur_id isn't initialized
>> - id = (unsigned int)atomic_inc_return(&cur_id);
>>
>> Please note that id is int. Is it expected that cur_id could
>> overflow?
>> Is there a maximum limit for this value?
>>
>
> Yes, we do expect cur_id to eventually overflow (once you have created
> 2 billion RPC client instances), however the atomic increment
> operations are expected to handle this correctly according to the
> maintainers (I already asked them in a different context). Furthermore,
> the code itself doesn't care about strict sequentiality. All it wants
> from the counter is uniqueness, with that uniqueness condition actually
> being enforced by the subsequent debugfs_create_file() call.
>
> IOW: I don't think this is a real problem.
>
Great. Thank you for a detailed explanation.
-- Shuah