Re: [PATCH] RDMA/core: Remove rcu attr for uverbs_api_ioctl_method.handler

[Yonghong Song <yhs@meta.com>]

On 10/25/22 9:30 AM, Jason Gunthorpe wrote:
> On Tue, Oct 25, 2022 at 09:29:09AM -0700, Paul E. McKenney wrote:
>> On Tue, Oct 25, 2022 at 01:00:02PM -0300, Jason Gunthorpe wrote:
>>> On Tue, Oct 25, 2022 at 08:24:20AM -0700, Yonghong Song wrote:
>>>> The current uverbs_api_ioctl_method definition:
>>>>    struct uverbs_api_ioctl_method {
>>>>          int(__rcu *handler)(struct uverbs_attr_bundle *attrs);
>>>>          DECLARE_BITMAP(attr_mandatory, UVERBS_API_ATTR_BKEY_LEN);
>>>>          ...
>>>>    };
>>>> The struct member 'handler' is marked with __rcu. But unless
>>>> the function body pointed by 'handler' is changing (e.g., jited)
>>>> during runtime, there is no need with __rcu.
>>>
>>> Huh? This is a sparse marker, it says that the pointer must always be
>>> loaded with rcu_dereference
>>>
>>> It has nothing to do with JIT, this patch is not correct
>>
>> OK, I will bite...
>>
>> This is a pointer to a function.  Given that this function's code is
>> generated at compile time, what sequence of changes is rcu_dereference()
>> protecting against?
> 
> Module unload. We set the value to NULL and then synchronize_rcu
> before unloading the code it points at.

Okay. Thanks for explanation. I forgot the module unload case.
I will look at compiler side then.

> 
> Jason