RE: Re:Re: Re:RE: Why has to offer "/dev/drv/render128" fd when running ROCm demo?

["Russell, Kent" <Kent.Russell@xxxxxxx>]

[AMD Official Use Only - General]

That flow was removed back in 5.13 timeframe, with this patch:

 

commit f45e6b9d03ff767ff946b6fce7ed3b004ab43ba4

Author: Felix Kuehling Felix.Kuehling@xxxxxxx

Date:   Wed Apr 7 17:30:05 2021 -0400

 

    drm/amdkfd: Remove legacy code not acquiring VMs

 

    ROCm user mode has acquired VMs from DRM file descriptors for as long

    as it supported the upstream KFD. Legacy code to support older versions

    of ROCm is not needed any more.

 

    Signed-off-by: Felix Kuehling Felix.Kuehling@xxxxxxx

    Reviewed-by: Philip Yang Philip.Yang@xxxxxxx

 

It was there from back in KFD’s early days, and has no longer been part of the upstream code since the 5.13 kernel. With KFD merging into amdgpu, and with ROCm using VMs from DRM FDs, there is no use case anymore where we would not use a DRM FD to get the VM.

 

Kent

 

From: amd-gfx <amd-gfx-bounces@xxxxxxxxxxxxxxxxxxxxx> On Behalf Of ???
Sent: Saturday, March 9, 2024 2:28 AM
To: Deucher, Alexander <Alexander.Deucher@xxxxxxx>
Cc: amd-gfx@xxxxxxxxxxxxxxxxxxxxx; Yat Sin, David <David.YatSin@xxxxxxx>; Kuehling, Felix <Felix.Kuehling@xxxxxxx>
Subject: Re:Re: Re:RE: Why has to offer "/dev/drv/render128" fd when running ROCm demo?

 

Hi alex/guys:

   No such situation exists.  there is always a rendernode.

=======>

if so, why the logic in function "kfd_process_device_init_vm" has a situation that in consideration of the "drm render file" not exsists? and even im "kfd_bind_process_to_device" therea are lazy vm conetxt binding " err = kfd_process_device_init_vm(pdd, NULL);" which pass NULL pointer as  drm file directly ?   if so , can the another path on scenario drm_file empty "amdgpu_amdkfd_gpuvm_create_process_vm"  call is redundant? 

 

Thank you!

774 int kfd_process_device_init_vm(struct kfd_process_device *pdd,

 775                         ¦      struct file *drm_file)

 776 {       

 777         struct kfd_process *p;

 778         struct kfd_dev *dev;

 779         int ret;

 780         

 781         if (pdd->vm)

 782                 return drm_file ? -EBUSY : 0;

 783         

 784         p = pdd->process;

 785         dev = pdd->dev;

 786         

 787         if (drm_file)

 788                 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(

 789                         dev->kgd, drm_file, p->pasid,

 790                         &pdd->vm, &p->kgd_process_info, &p->ef);

 791         else

 792                 ret = amdgpu_amdkfd_gpuvm_create_process_vm(dev->kgd, p->pasid,

 793                         &pdd->vm, &p->kgd_process_info, &p->ef);

 794         if (ret) {

 795                 pr_err("Failed to create process VM object\n");

 796                 return ret;

 797         }

 

 

 

 

 

 

At 2024-03-06 22:45:13, "Deucher, Alexander" <Alexander.Deucher@xxxxxxx> wrote:

[Public]

 

No such situation exists.  there is always a rendernode.

 

Alex

 

---

From: 曹子龙 <gainerying@xxxxxxx>
Sent: Wednesday, March 6, 2024 12:25 AM
To: amd-gfx@xxxxxxxxxxxxxxxxxxxxx <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>
Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Yat Sin, David <David.YatSin@xxxxxxx>; Kuehling, Felix <Felix.Kuehling@xxxxxxx>
Subject: Re:RE: Why has to offer "/dev/drv/render128" fd when running ROCm demo?

 

Hi felix:

   Thanks for your kindly help.

 

i still has a question, if on a pure-compute platform, which has no /dev/dri/render node, but only with a single /dev/kfd node, how the compue scenario works since that wei have no "render" fd exists?

is such platfom(without render node) exists? 

 

 thanks for your kindly help.!

 

BRs

zlcao.

 

 

 

 

At 2024-03-06 04:43:30, "Kuehling, Felix" <Felix.Kuehling@xxxxxxx> wrote:

[AMD Official Use Only - General]

 

I already answered this question in a reply to another email that was addressed to me:

 

The render nodes are used for CPU mapping of VRAM with mmap calls and an offset that identifies the BO. The render node also creates the GPU virtual address space that is used by KFD to create the GPU memory mappings. Applications that use both graphics and compute can share the same GPU virtual address space in this way.

 

Zlcao, in the future, this type of question may be better addressed to a mailing list such as amd-gfx@xxxxxxxxxxxxxxxxxxxxx, instead of writing separately to different maintainers.

 

Regards,
  Felix

 

 

From: Deucher, Alexander <Alexander.Deucher@xxxxxxx>
Sent: Tuesday, March 5, 2024 3:17 PM
To: 曹子龙 <gainerying@xxxxxxx>; Kuehling, Felix <Felix.Kuehling@xxxxxxx>; Yat Sin, David <David.YatSin@xxxxxxx>
Subject: Re: Why has to offer "/dev/drv/render128" fd when running ROCm demo?

 

[AMD Official Use Only - General]

 

+ Felix, David to comment.

---

From: 曹子龙 <gainerying@xxxxxxx>
Sent: Tuesday, March 5, 2024 8:46 AM
To: Deucher, Alexander <Alexander.Deucher@xxxxxxx>
Subject: Why has to offer "/dev/drv/render128" fd when running ROCm demo?

 

Hi alexander.:

      sorry for bother. but i really need some help to deal with my puzzle. 

i am a freshman to AMD GPU Driver, NOW i am trying to write a simple demo to use "/dev/kfd" and do some  vram alloc/free tests,  but i found you must do the right "AMDKFD_IOC_ACQUIRE_VM" ioctl comamnd before you try GPU to do vram allocation and other things.  from the kfd driver code,the pre-condition to do the right "AMDKFD_IOC_ACQUIRE_VM" is to offer a "/dev/drv/render128" fd  to the parameters.

 

so, why need to do this? kfd is used for compute, but "/dev/dri/render128" is specify to gfx usaged, why must open the "/dev/dri/render128" during the KFD compute scenario?

thanks for your kindly help!

 

BRs

zlcao.