On 6/15/23 13:40, Christoph Hellwig wrote:
> On Thu, Jun 15, 2023 at 09:10:28AM +0900, Damien Le Moal wrote:
>>> Here's what commit e27829dc92e5 ("scsi: serialize ->rescan against
>>> ->remove", written by Christoph Hellwig) says:
>>>
>>> Lock the device embedded in the scsi_device to protect against
>>> concurrent calls to ->remove.
>>>
>>> That's the commit which added the device_lock() call.
>>
>> Thanks for the information.
>>
>> +Christoph
>>
>> Why is adding the device_lock() needed ? We could just do a
>> scsi_device_get()+scsi_device_put() to serialize against remove. No ?
>
> No. scsi_device_get just increments a reference count, and thus
> prevents ->release from beeing called. ->remove is not in any way
> affected by the refcount.
What ->remove cb are you talking about ? The gendev one ?
I am trying to understand why the use of device_lock() helps in any way given
that this is not used by any other functions in scsi. And given that
scsi_rescan_device() should always be called with a ref on the scsi device (and
so on the gendev as well) held, why would this function be racy with device remove ?
Note that I did find a couple of places where scsi_rescan_device() seems to not
be called with a reference to the scsi dev held, e.g. store_rescan_field() and
store_state_field().
--
Damien Le Moal
Western Digital Research
