[PATCH v2] drm/amdgpu: fix the memory corruption on S3

[Huang Rui]

On Thu, Jun 29, 2017 at 04:16:53PM +0800, Christian König wrote:
> Am 29.06.2017 um 10:09 schrieb Huang Rui:
> > psp->cmd will be used on resume phase, so we can not free it on hw_init.
> > Otherwise, a memory corruption will be triggered.
> >
> > Signed-off-by: Huang Rui <ray.huang at amd.com>
> > ---
> >
> > V1 -> V2:
> > - remove "cmd" variable.
> > - fix typo of check.
> >
> > Alex, Christian,
> >
> > This is the final fix for vega10 S3. The random memory corruption issue is
> root
> > caused.
> >
> > Thanks,
> > Ray
> >
> > ---
> >   drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 17 +++++++++--------
> >   1 file changed, 9 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c b/drivers/gpu/drm/amd/
> amdgpu/amdgpu_psp.c
> > index 5bed483..711476792 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c
> > @@ -330,14 +330,11 @@ static int psp_load_fw(struct amdgpu_device *adev)
> >   {
> >        int ret;
> >        struct psp_context *psp = &adev->psp;
> > -     struct psp_gfx_cmd_resp *cmd;
> >  
> > -     cmd = kzalloc(sizeof(struct psp_gfx_cmd_resp), GFP_KERNEL);
> > -     if (!cmd)
> > +     psp->cmd = kzalloc(sizeof(struct psp_gfx_cmd_resp), GFP_KERNEL);
> > +     if (!psp->cmd)
> >                return -ENOMEM;
> >  
> > -     psp->cmd = cmd;
> > -
> >        ret = amdgpu_bo_create_kernel(adev, PSP_1_MEG, PSP_1_MEG,
> >                                      AMDGPU_GEM_DOMAIN_GTT,
> >                                      &psp->fw_pri_bo,
> > @@ -376,8 +373,6 @@ static int psp_load_fw(struct amdgpu_device *adev)
> >        if (ret)
> >                goto failed_mem;
> >  
> > -     kfree(cmd);
> > -
> >        return 0;
> >  
> >   failed_mem:
> > @@ -387,7 +382,8 @@ static int psp_load_fw(struct amdgpu_device *adev)
> >        amdgpu_bo_free_kernel(&psp->fw_pri_bo,
> >                              &psp->fw_pri_mc_addr, &psp->fw_pri_buf);
> >   failed:
> > -     kfree(cmd);
> > +     kfree(psp->cmd);
> > +     psp->cmd = NULL;
> >        return ret;
> >   }
> >  
> > @@ -447,6 +443,11 @@ static int psp_hw_fini(void *handle)
> >                amdgpu_bo_free_kernel(&psp->fence_buf_bo,
> >                                      &psp->fence_buf_mc_addr, &psp->
> fence_buf);
> >  
> > +     if (psp->cmd) {
> 
> As Michel noted as well please drop this extra check, kfree(NULL) is
> perfectly save.
> 
> With that fixed the patch is Reviewed-by: Christian König
> <christian.koenig at amd.com> for now, but I still think we could do better
> by only allocating the temporary command buffer when it is needed.
> 

Thanks. This is the quick fix for release. You know, it was a tragedy till
I found the root cause for S3 suspend/resume and make it stable, now it's
able to enter S3 more than 30+ cycles and never crash. 

I am planning to refine the psp codes, any suggestions are warm for me. I
will refer the comments such as fence and "temporary command buffter" to
modify it in following days. :-)

Thanks,
Ray