RE: [PATCH 2/2 v3] drm/exynos: added userptr feature.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Jerome,

Thank you again.

> -----Original Message-----
> From: Jerome Glisse [mailto:j.glisse@xxxxxxxxx]
> Sent: Thursday, May 10, 2012 3:33 AM
> To: Inki Dae
> Cc: airlied@xxxxxxxx; dri-devel@xxxxxxxxxxxxxxxxxxxxx;
> kyungmin.park@xxxxxxxxxxx; sw0312.kim@xxxxxxxxxxx; linux-mm@xxxxxxxxx
> Subject: Re: [PATCH 2/2 v3] drm/exynos: added userptr feature.
> 
> On Wed, May 9, 2012 at 10:45 AM, Jerome Glisse <j.glisse@xxxxxxxxx> wrote:
> > On Wed, May 9, 2012 at 2:17 AM, Inki Dae <inki.dae@xxxxxxxxxxx> wrote:
> >> this feature is used to import user space region allocated by malloc()
> or
> >> mmaped into a gem. and to guarantee the pages to user space not to be
> >> swapped out, the VMAs within the user space would be locked and then
> unlocked
> >> when the pages are released.
> >>
> >> but this lock might result in significant degradation of system
> performance
> >> because the pages couldn't be swapped out so we limit user-desired
> userptr
> >> size to pre-defined.
> >>
> >> Signed-off-by: Inki Dae <inki.dae@xxxxxxxxxxx>
> >> Signed-off-by: Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
> >
> >
> > Again i would like feedback from mm people (adding cc). I am not sure
> > locking the vma is the right anwser as i said in my previous mail,
> > userspace can munlock it in your back, maybe VM_RESERVED is better.
> > Anyway even not considering that you don't check at all that process
> > don't go over the limit of locked page see mm/mlock.c RLIMIT_MEMLOCK
> > for how it's done. Also you mlock complete vma but the userptr you get
> > might be inside say 16M vma and you only care about 1M of userptr, if
> > you mark the whole vma as locked than anytime a new page is fault in
> > the vma else where than in the buffer you are interested then it got
> > allocated for ever until the gem buffer is destroy, i am not sure of
> > what happen to the vma on next malloc if it grows or not (i would
> > think it won't grow at it would have different flags than new
> > anonymous memory).
> >
> > The whole business of directly using malloced memory for gpu is fishy
> > and i would really like to get it right rather than relying on never
> > hitting strange things like page migration, vma merging, or worse
> > things like over locking pages and stealing memory.
> >
> > Cheers,
> > Jerome
> 
> I had a lengthy discussion with mm people (thx a lot for that). I
> think we should split 2 different use case. The zero-copy upload case
> ie :
> app:
>     ptr = malloc()
>     ...
>     glTex/VBO/UBO/...(ptr)
>     free(ptr) or reuse it for other things
> For which i guess you want to avoid having to do a memcpy inside the
> gl library (could be anything else than gl that have same useage
> pattern).
> 

Right, in this case, we are using the userptr feature as pixman and evas
backend to use 2d accelerator.

> ie after the upload happen you don't care about those page they can
> removed from the vma or marked as cow so that anything messing with
> those page after the upload won't change what you uploaded. Of course

I'm not sure that I understood your mentions but could the pages be removed
from vma with VM_LOCKED or VM_RESERVED? once glTex/VBO/UBO/..., the VMAs to
user space would be locked. if cpu accessed significant part of all the
pages in user mode then pages to the part would be allocated by page fault
handler, after that, through userptr, the VMAs to user address space would
be locked(at this time, the remaining pages would be allocated also by
get_user_pages by calling page fault handler) I'd be glad to give me any
comments and advices if there is my missing point.

> this is assuming that the tlb cost of doing such thing is smaller than
> the cost of memcpy the data.
> 

yes, in our test case, the tlb cost(incurred by tlb miss) was smaller than
the cost of memcpy also cpu usage. of course, this would be depended on gpu
performance.

> Two way to do that, either you assume app can't not read back data
> after gl can and you do an unmap_mapping_range (make sure you only
> unmap fully covered page and that you copy non fully covered page) or
> you want to allow userspace to still read data or possibly overwrite
> them
> 
> Second use case is something more like for the opencl case of
> CL_MEM_USE_HOST_PTR, in which you want to use the same page in the gpu
> and keep the userspace vma pointing to those page. I think the
> agreement on this case is that there is no way right now to do it
> sanely inside linux kernel. mlocking will need proper accounting
> against rtlimit but this limit might be low. Also the fork case might
> be problematic.
> 
> For the fork case the memory is anonymous so it should be COWed in the
> fork child but relative to cl context that means the child could not
> use the cl context with that memory or at least if the child write to
> this memory the cl will not see those change. I guess the answer to
> that one is that you really need to use the cl api to read the object
> or get proper ptr to read it.
> 
> Anyway in all case, implementing this userptr thing need a lot more
> code. You have to check to that the vma you are trying to use is
> anonymous and only handle this case and fallback to alloc new page and
> copy otherwise..
> 

I'd like to say thank you again you gave me comments and advices in detail.
there may be my missing points but I will check it again.

Thanks,
Inki Dae

> Cheers,
> Jerome

_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/dri-devel


[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux