Re: [PATCH] drm/amdgpu: fix modprobe failure of the 2nd GPU when GDDR6 training enabled

["Yin, Tianci (Rico)" <Tianci.Yin@xxxxxxx>]

[AMD Official Use Only - Internal Distribution Only]

Thanks Alex and Christian!

Hi Christian,

On ASICs with gmc v10, I find amdgpu_bo_late_init() is not invoked, so stolen memory never get free, on other ASICs with gmc v9/v8/v7/v6, stolen memory was freed in gmc_v*_0_late_init(). I don't know why, are there some special reasons or just missed by coding?

Thanks!

Rico

---

From: Christian König <ckoenig.leichtzumerken@xxxxxxxxx>
Sent: Wednesday, January 8, 2020 23:04
To: Alex Deucher <alexdeucher@xxxxxxxxx>; Koenig, Christian <Christian.Koenig@xxxxxxx>
Cc: Long, Gang <Gang.Long@xxxxxxx>; Wang, Kevin(Yang) <Kevin1.Wang@xxxxxxx>; Xu, Feifei <Feifei.Xu@xxxxxxx>; Yin, Tianci (Rico) <Tianci.Yin@xxxxxxx>; amd-gfx list <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>; Tuikov, Luben <Luben.Tuikov@xxxxxxx>; Deucher, Alexander <Alexander.Deucher@xxxxxxx>; Yuan, Xiaojie <Xiaojie.Yuan@xxxxxxx>; Zhang, Hawking <Hawking.Zhang@xxxxxxx>
Subject: Re: [PATCH] drm/amdgpu: fix modprobe failure of the 2nd GPU when GDDR6 training enabled

 

Am 08.01.20 um 15:49 schrieb Alex Deucher:
> On Wed, Jan 8, 2020 at 7:52 AM Christian König <christian.koenig@xxxxxxx> wrote:
>> Am 08.01.20 um 13:36 schrieb Tianci Yin:
>>> From: "Tianci.Yin" <tianci.yin@xxxxxxx>
>>>
>>> [why]
>>> In dual GPUs scenario, stolen_size is assigned to zero on the 2nd GPU,
>>> then the bottom region of VRAM was allocated as GTT, unfortunately
>>> a small region of bottom VRAM was encroached by UMC firmware during
>>> GDDR6 BIST training, this cause pagefault.
>> What I'm missing here is why is the stolen size zero on the 2nd GPU?
>>
>> Maybe we need to read the stolen size after posting the GPU instead?
> There is no stolen memory on secondary GPUs because there is no pre-OS
> console using that memory.

Ah! Yeah that makes sense.

But in this case I would say we should change the patch like the following:

adev->gmc.stolen_size = min(adev->gmc.stolen_size,
AMDGPU_STOLEN_VGA_DEFAULT_SIZE);

And in amdgpu_ttm_late_init():

/* Can't free the stolen VGA memory when it might be used for memory
training again. */
if (!adev->fw_vram_usage.mem_train_support)
     amdgpu_bo_free_kernel(....


Regards,
Christian.


>
> Alex
>
>> Regards,
>> Christian.
>>
>>> [how]
>>> Forcing stolen_size to 3MB, then the bottom region of VRAM was
>>> allocated as stolen memory, GTT corruption avoid.
>>> The stolen memory of the 2nd GPU will be free in late_init phase,
>>> no memory wasted.
>>>
>>> Change-Id: Icd0ad7de41333282949bb1e3e676c6c307ddd081
>>> Signed-off-by: Tianci.Yin <tianci.yin@xxxxxxx>
>>> ---
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  6 ++++++
>>>    drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  | 21 +++++++++++++++++++++
>>>    2 files changed, 27 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
>>> index c91dd602d5f1..440b793316df 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
>>> @@ -60,6 +60,11 @@
>>>     */
>>>    #define AMDGPU_GMC_FAULT_TIMEOUT    5000ULL
>>>
>>> +/*
>>> + * Default stolen memory size, 1024 * 768 * 4
>>> + */
>>> +#define AMDGPU_STOLEN_VGA_DEFAULT_SIZE       0x300000
>>> +
>>>    struct firmware;
>>>
>>>    /*
>>> @@ -192,6 +197,7 @@ struct amdgpu_gmc {
>>>        uint32_t                srbm_soft_reset;
>>>        bool                    prt_warning;
>>>        uint64_t                stolen_size;
>>> +     bool                    stolen_temp_reserved;
>>>        /* apertures */
>>>        u64                     shared_aperture_start;
>>>        u64                     shared_aperture_end;
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>>> index 7dc8c068c62a..0c96b67d6ca7 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>>> @@ -566,6 +566,11 @@ static int gmc_v10_0_late_init(void *handle)
>>>        struct amdgpu_device *adev = (struct amdgpu_device *)handle;
>>>        int r;
>>>
>>> +     if (adev->gmc.stolen_temp_reserved) {
>>> +             amdgpu_bo_late_init(adev);
>>> +             adev->gmc.stolen_temp_reserved = false;
>>> +     }
>>> +
>>>        r = amdgpu_gmc_allocate_vm_inv_eng(adev);
>>>        if (r)
>>>                return r;
>>> @@ -756,6 +761,22 @@ static int gmc_v10_0_sw_init(void *handle)
>>>                return r;
>>>
>>>        adev->gmc.stolen_size = gmc_v10_0_get_vbios_fb_size(adev);
>>> +     /*
>>> +      * In dual GPUs scenario, stolen_size is assigned to zero on the 2nd GPU,
>>> +      * then the bottom region of VRAM was allocated as GTT, unfortunately
>>> +      * a small region of bottom VRAM was encroached by UMC firmware during
>>> +      * GDDR6 BIST training, this cause pagefault.
>>> +      * The page fault can be fixed by forcing stolen_size to 3MB, then the bottom
>>> +      * region of VRAM was allocated as stolen memory, GTT corruption avoid.
>>> +      * The stolen memory of the 2nd GPU will be free in late_init phase,
>>> +      * no memory wasted.
>>> +      */
>>> +     if (adev->fw_vram_usage.mem_train_support &&
>>> +             adev->gmc.stolen_size == 0) {
>>> +             adev->gmc.stolen_size = AMDGPU_STOLEN_VGA_DEFAULT_SIZE;
>>> +             adev->gmc.stolen_temp_reserved = true;
>>> +     } else
>>> +             adev->gmc.stolen_temp_reserved = false;
>>>
>>>        /* Memory manager */
>>>        r = amdgpu_bo_init(adev);
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx@xxxxxxxxxxxxxxxxxxxxx
>> https://nam11.safelinks.protection.outlook.com/?url="">
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@xxxxxxxxxxxxxxxxxxxxx
> https://nam11.safelinks.protection.outlook.com/?url="">

_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx