Re: [RFC PATCH] mm, pmem, xfs: Introduce MF_MEM_REMOVE for unbind

[Shiyang Ruan <ruansy.fnst@fujitsu.com>]

在 2022/4/11 15:06, Christoph Hellwig 写道:
> On Mon, Apr 11, 2022 at 01:16:23AM +0800, Shiyang Ruan wrote:
>> diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
>> index bd502957cfdf..72d9e69aea98 100644
>> --- a/drivers/nvdimm/pmem.c
>> +++ b/drivers/nvdimm/pmem.c
>> @@ -359,7 +359,6 @@ static void pmem_release_disk(void *__pmem)
>>   	struct pmem_device *pmem = __pmem;
>>   
>>   	dax_remove_host(pmem->disk);
>> -	kill_dax(pmem->dax_dev);
>>   	put_dax(pmem->dax_dev);
>>   	del_gendisk(pmem->disk);
>>   
>> @@ -597,6 +596,8 @@ static void nd_pmem_remove(struct device *dev)
>>   		pmem->bb_state = NULL;
>>   	}
>>   	nvdimm_flush(to_nd_region(dev->parent), NULL);
>> +
>> +	kill_dax(pmem->dax_dev);
> 
> I think the put_dax will have to move as well.

After reading the implementation of 'devm_add_action_or_reset()', I 
think there is no need to move kill_dax() and put_dax() into ->remove().

In unbind, it will call both drv->remove() and devres_release_all(). 
The action, pmem_release_disk(), added in devm_add_action_or_reset() 
will be execute in devres_release_all().  So, during the unbind process, 
{kill,put}_dax() will finally be called to notify the REMOVE signal.

In addition, if devm_add_action_or_reset() fails in pmem_attach_disk(), 
pmem_release_disk() will be called to cleanup the pmem->dax_dev.


--
Thanks,
Ruan.

> 
> This part should probably also be a separate, well-documented
> cleanup patch.