Re: [PATCH v2] drm/ttm: Silence randstruct warning about casting struct file

["Christian König" <christian.koenig@amd.com>]

On 5/2/25 07:33, Al Viro wrote:
> On Thu, May 01, 2025 at 09:52:08PM -0700, Matthew Brost wrote:
>> On Fri, May 02, 2025 at 05:31:49AM +0100, Al Viro wrote:
>>> On Thu, May 01, 2025 at 09:26:25PM -0700, Matthew Brost wrote:
> 
>>> And what is the lifecycle of that thing?  E.g. what is guaranteed about
>>> ttm_backup_fini() vs. functions accessing the damn thing?  Are they
>>> serialized on something/tied to lifecycle stages of struct ttm_tt?
>>
>> I believe the life cycle is when ttm_tt is destroyed or api allows
>> overriding the old backup with a new one (currently unused).
> 
> Umm...  So can ttm_tt_setup_backup() be called in the middle of
> e.g. ttm_backup_drop() or ttm_backup_{copy,backup}_page(), etc.?
> 
> I mean, if they had been called by ttm_backup.c internals, it would
> be an internal business of specific implementation, with all
> serialization, etc. warranties being its responsibility;
> but if it's called by other code that is supposed to be isolated
> from details of what ->backup is pointing to...
> 
> Sorry for asking dumb questions, but I hadn't seen the original
> threads.  Basically, what prevents the underlying shmem file getting
> torn apart while another operation is using it?  It might very well
> be simple, but I had enough "it's because of... oh, bugger" moments
> on the receiving end of such questions...

It's the outside logic which makes sure that the backup structure stays around as long as the BO or the device which needs it is around.

But putting that aside I was not very keen about the whole idea of never defining the ttm_backup structure and just casting it to a file in the backend either.

So I would just completely nuke that unnecessary abstraction and just use a pointer to a file all around.

Regards,
Christian.