Re: ttm_mem_global

["Thomas Hellström" <thomas@shipmail.org>]

Jerome Glisse wrote:
> On Tue, 2009-07-28 at 20:55 +0200, Thomas Hellström wrote:
>   
>> Jerome Glisse skrev:
>>     
>>> On Wed, 2009-07-22 at 10:37 +0200, Thomas Hellström wrote:
>>>   
>>>       
>>>> TTM has a device struct per device and an optional global struct that is 
>>>> common for all devices and intended to be per subsystem.
>>>>
>>>> The only subsystem currently having a global structure is the memory 
>>>> accounting subsystem:
>>>> struct ttm_mem_global
>>>>     
>>>>         
>>> Thomas i don't think the way we init ttm_mem_global today make
>>> it follow the 1 struct ttm_mem_global for everyone. I think it
>>> should be initialized and refcounted by device struct.
>>>
>>> So on first device creation a ttm_mem_global is created and
>>> then anytime a new device is created the refcount of ttm_mem_global
>>> is increased. 
>>>       
>> Jerome,
>> This is exactly what the current code intends to do.
>>
>> Are you seeing something different?
>>     
>
> I definitly don't see that :) In radeon we do create a structure
> which hold the ttm_mem_global struct so it's not shared at all
> it got inited & destroyed along the driver. This is why i think
> it's better to remove the driver initialization and let bo_device
> init path take care of initializing one and only one object which
> can be shared by multiple driverttm_mem_global_inits.
>
>   
Which radeon struct is holding the ttm_mem_global struct?

The radeon code looks very similar to the openchrome code in which the 
struct ttm_mem_global is allocated at ttm_global.c, line 74 and freed at 
ttm_global.c, line 108 when its refcount has reached zero.

So the device holds a struct ttm_global_reference that *only points* to 
the global item, and which is destroyed on device takedown. If there are 
more than one device pointing to the mem_global object, it won't get 
destroyed.

So the code should be working perfectly fine unless there is a bug.

> So what i propose is remove mem_glob parameter from :
> ttm_bo_device_init, add a call to ttm_mem_global_init in
> ttm_bo_device_init 

Nope, The ttm_mem_global object is used by other  ttm subsystems 
(fencing, user-space objects),
so that can't be done.

> and add some static refcount in ttm_memory.c
> if refcount = 0 then ttm_mem_global_init create a ttm_mem_global
> struct and initialize things, if refcount > 0 then it gives
> back the already initialized ttm_mem_global.
>
>   

This is exactly what ttm_global was created to do, and what it hopefully 
does. If you create two radeon devices the ttm_mem_global object should 
be the same, even though the global references pointing to it are of 
course different. Have you actually tried this?

/Thomas

> Of course we unref with ttm_mem_global_release and destroy
> once refcount reach 0.
>
> Cheers,
> Jerome
>
>