Why there is no GPU swap support in linux based systems?
(The things below are mainly related to arm mali systems) Hi. Actually i want to know - why there is no GPU Swap support in Linux based systems? It seems that the GPU works with graphical contexts in the same fashion as CPU handle things. I mean there are page faults, GPU have its own MMU, the memory used by GPU devices allocated via trivial page_alloc mechanism. You would say that performance is a reason, but there are some swap backends like zram: they use in-memory compression instead of disk swapping. Moreover, i think there would be a GPU data which could be swapped without significant drawbacks, because it accessed very rare. Thank you.
On Tue, Feb 09, 2016 at 09:12:23PM +0300, alexhoppus wrote:
(The things below are mainly related to arm mali systems)
Hi. Actually i want to know - why there is no GPU Swap support in Linux based systems? It seems that the GPU works with graphical contexts in the same fashion as CPU handle things. I mean there are page faults, GPU have its own MMU, the memory used by GPU devices allocated via trivial page_alloc mechanism. You would say that performance is a reason, but there are some swap backends like zram: they use in-memory compression instead of disk swapping. Moreover, i think there would be a GPU data which could be swapped without significant drawbacks, because it accessed very rare.
That's not how GPU devices work. But if you think you could help implement this, I'm sure the graphic developers are always welcome to working patches being sent to them for review. best of luck, greg k-h
On 09.02.2016 22:33, Greg KH wrote:
On Tue, Feb 09, 2016 at 09:12:23PM +0300, alexhoppus wrote:
(The things below are mainly related to arm mali systems)
Hi. Actually i want to know - why there is no GPU Swap support in Linux based systems? It seems that the GPU works with graphical contexts in the same fashion as CPU handle things. I mean there are page faults, GPU have its own MMU, the memory used by GPU devices allocated via trivial page_alloc mechanism. You would say that performance is a reason, but there are some swap backends like zram: they use in-memory compression instead of disk swapping. Moreover, i think there would be a GPU data which could be swapped without significant drawbacks, because it accessed very rare. That's not how GPU devices work. But if you think you could help implement this, I'm sure the graphic developers are always welcome to working patches being sent to them for review.
best of luck,
greg k-h
Could you (anyone), please, explain what exactly is wrong with this? I'am not talking about discrete video cards. As i understand, in this case video memory is a subset of pages from main memory. Also i will highlight that i'am talking about swapping out some GPU related data f.e. textures, for tasks which are not active at the moment (background graphical contexts). As i see both CPU and GPU can handle page faults for graphical memory pages and we can establish apropriate CPU and GPU mappings for this pages. So what is the problem with this then? Thanks for response.
participants (2)
-
alexhoppus -
Greg KH