Re: [RFC PATCH] fpga: remove module reference counting from core components

[Marco Pagani]

On 2023-11-08 16:52, Xu Yilun wrote:
> On Fri, Nov 03, 2023 at 09:31:02PM +0100, Marco Pagani wrote:
>>
>>
>> On 2023-10-30 09:32, Xu Yilun wrote:
>>> On Fri, Oct 27, 2023 at 05:29:27PM +0200, Marco Pagani wrote:
>>>> Remove unnecessary module reference counting from the core components
>>>> of the subsystem. Low-level driver modules cannot be removed before
>>>> core modules since they use their exported symbols.
>>>
>>> Could you help show the code for this conclusion?
>>>
>>> This is different from what I remember, a module cannot be removed when
>>> its exported symbols are being used by other modules. IOW, the core
>>> modules cannot be removed when there exist related low-level driver
>>> modules. But the low-level driver modules could be removed freely
>>> without other protecting mechanism.
>>>
>>
>> My understanding was that we wanted to remove module reference counting
>> from the fpga core and ease it from the responsibility of preventing
>> low-level driver modules from being unloaded. 
> 
> FPGA core needs to prevent low-level driver module unloading sometimes,
> e.g. when region reprograming is in progress. That's why we get fpga
> region driver modules & bridge modules in fpga_region_program_fpga().
> 
> But we try best to get them only necessary. Blindly geting them all the
> time results in no way to unload all modules (core & low level modules).
> 
>>
>> If we want to keep reference counting in the fpga core, we could add a
>> struct module *owner field in the struct fpga_manager_ops (and others
>> core *_ops) so that the low-level driver can set it to THIS_MODULE.
>> In this way, we can later use it in fpga_mgr_register() to bump up the
> 
> Yes, we should pass the module owner in fpga_mgr_register(), but could
> not bump up its refcount at once.
> 
>> refcount of the low-level driver module by calling
>> try_module_get(mgr->mops->owner) directly when it registers the manager.
>> Finally, fpga_mgr_unregister() would call module_put(mgr->mops->owner)
>> to allow unloading the low-level driver module.
> 
> As mentioned above, that makes problem. Most of the low level driver
> modules call fpga_mgr_unregister() on module_exit(), but bumping up
> their module refcount prevents module_exit() been executed. That came
> out to be a dead lock.
>

Initially, I considered calling try_module_get(mgr->mops->owner)
in fpga_mgr_get(). But then, the new kernel-doc description of
try_module_get() (1) made me question the safety of that approach.
My concern is that the low-level driver could be removed right when
someone is calling fpga_mgr_get() and hasn't yet reached
try_module_get(mgr->mops->owner). In that case, the struct mops
(along with the entire low-level driver module) and the manager dev
would "disappear" under the feet of fpga_mgr_get().

(1) 557aafac1153 ("kernel/module: add documentation for try_module_get()")
 
>>
>> In this way, it would no longer be necessary to call try_module_get()
>> in fpga_mrg_get() since we could use a kref (included in the struct
>> fpga_manager) to do refcounting for the in-kernel API users. Only when
>> the kref reaches zero fpga_mgr_unregister() would succeed and put the
>> low-level driver module.
>>
>> I think this approach would be safer since it would avoid the crash
>> that can currently happen if the low-level driver module is removed
>> right when executing try_module_get() in fpga_mrg_get(). The possible
>> caveat is that it would be required to call fpga_mgr_unregister()
>> before being able to remove the low-level driver module.
>>


[...]

Thanks,
Marco