[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <019a15d5-142f-4761-9408-58c103d3922b@amd.com>
Date: Tue, 1 Jul 2025 10:32:33 +0200
From: Christian König <christian.koenig@....com>
To: "Zhang, GuoQing (Sam)" <GuoQing.Zhang@....com>,
"rafael@...nel.org" <rafael@...nel.org>,
"len.brown@...el.com" <len.brown@...el.com>,
"pavel@...nel.org" <pavel@...nel.org>,
"Deucher, Alexander" <Alexander.Deucher@....com>,
"Limonciello, Mario" <Mario.Limonciello@....com>,
"Lazar, Lijo" <Lijo.Lazar@....com>
Cc: "Zhao, Victor" <Victor.Zhao@....com>, "Chang, HaiJun"
<HaiJun.Chang@....com>, "Ma, Qing (Mark)" <Qing.Ma@....com>,
"amd-gfx@...ts.freedesktop.org" <amd-gfx@...ts.freedesktop.org>,
"dri-devel@...ts.freedesktop.org" <dri-devel@...ts.freedesktop.org>,
"linux-pm@...r.kernel.org" <linux-pm@...r.kernel.org>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 3/3] drm/amdgpu: skip kfd resume_process for
dev_pm_ops.thaw()
On 01.07.25 10:03, Zhang, GuoQing (Sam) wrote:
>
> thaw() is called before writing the hiberation image to swap disk. See
> the doc here.
> https://github.com/torvalds/linux/blob/v6.14/Documentation/driver-api/pm/devices.rst?plain=1#L552 <https://github.com/torvalds/linux/blob/v6.14/Documentation/driver-api/pm/devices.rst?plain=1#L552>
>
> And amdgpu implemented thaw() callback by calling amdgpu_device_resume().
> https://github.com/torvalds/linux/blob/v6.14/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c#L2572 <https://github.com/torvalds/linux/blob/v6.14/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c#L2572>
>
> This patch is skip amdgpu_amdkfd_resume_process() call in thaw() during
> hibernation. it is not skipped in restore() during resume from
> hibernation when system boot again.
>
>
> I just found the following kernel doc. Thaw() is intended to resume the
> storage device for saving the hibernation image.
Ah, that makes much more sense.
> Our GPU is not involved
> in it, it is not necessary to resume our GPU in thaw().
> https://github.com/torvalds/linux/blob/v6.14/Documentation/power/pci.rst?plain=1#L588 <https://github.com/torvalds/linux/blob/v6.14/Documentation/power/pci.rst?plain=1#L588>
>
> So another implementation is to remove the amdgpu_device_resume() call
> in amdgpu_pmops_thaw(), and skip amdgpu_device_ip_suspend() call in
> amdgpu_pci_shutdown()for hibernation.
> Initial tests show it's working fine for hibernation successful case.
> Should I switch to this implementation?
No idea. Alex and the KFD guys need to take a look at that.
> But thaw() is also called to restore the GPU when hibernation is aborted
> due to some error in hibernation image creation stage. In this case,
> amdgpu_device_resume() is needed in thaw().
>
> So I need a method to check if hibernation is aborted or not to
> conditionally skip amdgpu_device_resume() in thaw(). Currently I don't
> know how to do this.
Yeah that approach here looks fishy to me, but I don't know how to properly fix it either.
@Alex any idea?
Regards,
Christian.
>
>
> Regards
> Sam
>
>
> On 2025/6/30 19:58, Christian König wrote:
>> On 30.06.25 12:41, Samuel Zhang wrote:
>>> The hibernation successful workflow:
>>> - prepare: evict VRAM and swapout GTT BOs
>>> - freeze
>>> - create the hibernation image in system memory
>>> - thaw: swapin and restore BOs
>> Why should a thaw happen here in between?
>>
>>> - complete
>>> - write hibernation image to disk
>>> - amdgpu_pci_shutdown
>>> - goto S5, turn off the system.
>>>
>>> During prepare stage of hibernation, VRAM and GTT BOs will be swapout to
>>> shmem. Then in thaw stage, all BOs will be swapin and restored.
>> That's not correct. This is done by the application starting again and not during thaw.
>>
>>> On server with 192GB VRAM * 8 dGPUs and 1.7TB system memory,
>>> the swapin and restore BOs takes too long (50 minutes) and it is not
>>> necessary since the follow-up stages does not use GPU.
>>>
>>> This patch is to skip BOs restore during thaw to reduce the hibernation
>>> time.
>> As far as I can see that doesn't make sense. The KFD processes need to be resumed here and that can't be skipped.
>>
>> Regards,
>> Christian.
>>
>>> Signed-off-by: Samuel Zhang <guoqing.zhang@....com>
>>> ---
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 2 ++
>>> 2 files changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> index a8f4697deb1b..b550d07190a2 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> @@ -5328,7 +5328,7 @@ int amdgpu_device_resume(struct drm_device *dev, bool notify_clients)
>>> amdgpu_virt_init_data_exchange(adev);
>>> amdgpu_virt_release_full_gpu(adev, true);
>>>
>>> - if (!adev->in_s0ix && !r && !adev->in_runpm)
>>> + if (!adev->in_s0ix && !r && !adev->in_runpm && !adev->in_s4)
>>> r = amdgpu_amdkfd_resume_process(adev);
>>> }
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> index 571b70da4562..23b76e8ac2fd 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> @@ -2734,7 +2734,9 @@ static int amdgpu_pmops_poweroff(struct device *dev)
>>> static int amdgpu_pmops_restore(struct device *dev)
>>> {
>>> struct drm_device *drm_dev = dev_get_drvdata(dev);
>>> + struct amdgpu_device *adev = drm_to_adev(drm_dev);
>>>
>>> + adev->in_s4 = false;
>>> return amdgpu_device_resume(drm_dev, true);
>>> }
>>>
>
Powered by blists - more mailing lists