[PATCH] drm/amd/amdgpu: Add gfxoff debugfs entry

[Tom St Denis]

On 2020-02-21 1:59 p.m., Alex Deucher wrote:
> On Fri, Feb 21, 2020 at 1:45 PM Tom St Denis <tom.stdenis at amd.com> wrote:
>> Write a 32-bit value of zero to disable GFXOFF and write a 32-bit
>> value of non-zero to enable GFXOFF.
>>
>> Signed-off-by: Tom St Denis <tom.stdenis at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 56 +++++++++++++++++++++
>>   1 file changed, 56 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>> index 7379910790c9..3bb74056b9d2 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>> @@ -842,6 +842,55 @@ static ssize_t amdgpu_debugfs_gpr_read(struct file *f, char __user *buf,
>>          return result;
>>   }
>>
>> +/**
>> + * amdgpu_debugfs_regs_gfxoff_write - Enable/disable GFXOFF
>> + *
>> + * @f: open file handle
>> + * @buf: User buffer to write data from
>> + * @size: Number of bytes to write
>> + * @pos:  Offset to seek to
>> + *
>> + * Write a 32-bit zero to disable or a 32-bit non-zero to enable
>> + */
>> +static ssize_t amdgpu_debugfs_gfxoff_write(struct file *f, const char __user *buf,
>> +                                        size_t size, loff_t *pos)
>> +{
>> +       struct amdgpu_device *adev = file_inode(f)->i_private;
>> +       ssize_t result = 0;
>> +       int r;
>> +
>> +       if (size & 0x3 || *pos & 0x3)
>> +               return -EINVAL;
>> +
>> +       r = pm_runtime_get_sync(adev->ddev->dev);
> Not really directly related to this patch, but If you are using umr
> for debugging, we should probably disable runtime pm, otherwise the
> entire GPU may be powered down between accesses.  There is already an
> interface to do that via the core kernel power subsystem in sysfs.
> E.g.,
> /sys/class/drm/card0/device/power/control
> /sys/class/drm/card0/device/power/runtime_status
> Something else to look at for umr.

We ran into something related to this for UVD/VCE access back in the 
day.  When powered down the MMIO registers are mirrored and accessible 
but while in transition they are not.  So we added a PG flag to the 
offset in the debugfs entry to flag when we need to take the pm mutex or 
not.




>
> We don't store the state for when we dynamically turn it off like this
> so if we get a GPU reset or a power management event (runtime pm or
> S3), GFXOFF will be re-enabled at that point.  This is just for
> debugging though so:
> Acked-by: Alex Deucher <alexander.deucher at amd.com>

Good to note.  Can I get a R-b from someone though so I can push this out?


Tom