[V2 1/1] drm/amdgpu/gfx8: add support kernel interface queue(KIQ)

[Yu, Xiangliang]

> -----Original Message-----
> From: Christian König [mailto:deathsimple at vodafone.de]
> Sent: Wednesday, January 04, 2017 10:13 PM
> To: Yu, Xiangliang <Xiangliang.Yu at amd.com>; Zhou, David(ChunMing)
> <David1.Zhou at amd.com>; amd-gfx at lists.freedesktop.org; Liu, Monk
> <Monk.Liu at amd.com>
> Subject: Re: [V2 1/1] drm/amdgpu/gfx8: add support kernel interface
> queue(KIQ)
> 
> Am 04.01.2017 um 14:26 schrieb Yu, Xiangliang:
> >>>> Am 28.12.2016 um 03:59 schrieb Yu, Xiangliang:
> >>>>>> -----Original Message-----
> >>>>>> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
> >>>>>> Behalf Of zhoucm1
> >>>>>> Sent: Wednesday, December 28, 2016 10:32 AM
> >>>>>> To: Yu, Xiangliang <Xiangliang.Yu at amd.com>; amd-
> >>>>>> gfx at lists.freedesktop.org
> >>>>>> Cc: Liu, Monk <Monk.Liu at amd.com>
> >>>>>> Subject: Re: [V2 1/1] drm/amdgpu/gfx8: add support kernel
> >>>>>> interface
> >>>>>> queue(KIQ)
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>> On 2016年12月28日 10:27, Yu, Xiangliang wrote:
> >>>>>>> +
> >>>>>>>> +	amdgpu_bo_kunmap(kiq->eop_obj);
> >>>>>>>> +
> >>>>>> I see it is in many places, which is don't need I think,
> >>>>>> amdgpu_bo_free_kernel will handle it if I'm correct.
> >>>>> Memory kmap should be free ASAP from memory management point.
> >>>> No, that is completely unnecessary as long as you don't run a 32bit
> >>>> system with tight address space.
> >>> We can't decide what OS to use.  I agree with monk's comment that
> >>> better
> >> to unmap it if cpu don't use.
> >>
> >> No it isn't. As I said this is only necessary on 32bit systems with
> >> tight address space. In other words we simply don't really support
> >> that configuration any more.
> >>
> >> If you don't need it then why do you want to kmap it in the first place?
> >>
> > In my patch, first get mqd address to fill out mqd context through kmap,
> then setup compute queue with kiq ring.
> >
> > And I see there are lot of similar usage in amdgpu, for example,
> gfx_v8_0_rlc_init function. Could you help to explain it?
> 
> Ah! So you only map it initially to fill in some data structure during
> initialization? Yeah, in this case just map the BO temporary, fill in the data and
> unmap it again.

Yes.

> 
> Sorry, judging from the comments I thought that you map/unmap the BO
> multiple times on each IOCTL. That is really something we should try to avoid.
> 
> Just make sure that you don't store the pointer returned by kmap anywhere,
> it is only valid temporarily.

Got it. Thanks!