[PATCH v2] drm/amdgpu: fix the ib test hang when gfx is in "idle" state
S, Shirish
sshankar at amd.com
Tue Apr 24 03:44:27 UTC 2018
On 4/24/2018 8:19 AM, Huang Rui wrote:
> "aaabaf4 drm/amdgpu: defer test IBs on the rings at boot (V3)"
> Above patch defers the execution of gfx/compute ib tests. However, at that time,
> the gfx may already go into idle state. If "idle" gfx receives command
> submission, then will get hang in the system. And it still has issue to
> dynamically enable/disable gfxoff at runtime. So we have to use a workaround to
> skip the gfx/compute ib tests when gfx is already in "idle" state.
>
> Signed-off-by: Huang Rui <ray.huang at amd.com>
> Cc: Shirish S <shirish.s at amd.com>
> ---
>
> Changes from V1 -> V2:
> - Remove unused definitions of smu10_hwmgr.
> - Add WA descriptions into commit log and comments.
>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 ++
> drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 23 ++++++++++++++++++++++-
Is it not required for older asics of cz/st?
If not then please update the commit message accordingly.
Regards,
Shirish S
> drivers/gpu/drm/amd/powerplay/hwmgr/smu10_hwmgr.c | 22 ++--------------------
> 3 files changed, 26 insertions(+), 21 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 59df4b7..a0263b9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -905,6 +905,7 @@ struct amdgpu_gfx_funcs {
> void (*read_wave_vgprs)(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t thread, uint32_t start, uint32_t size, uint32_t *dst);
> void (*read_wave_sgprs)(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t start, uint32_t size, uint32_t *dst);
> void (*select_me_pipe_q)(struct amdgpu_device *adev, u32 me, u32 pipe, u32 queue);
> + bool (*is_gfx_on)(struct amdgpu_device *adev);
> };
>
> struct amdgpu_ngg_buf {
> @@ -1855,6 +1856,7 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
> #define amdgpu_gds_switch(adev, r, v, d, w, a) (adev)->gds.funcs->patch_gds_switch((r), (v), (d), (w), (a))
> #define amdgpu_psp_check_fw_loading_status(adev, i) (adev)->firmware.funcs->check_fw_loading_status((adev), (i))
> #define amdgpu_gfx_select_me_pipe_q(adev, me, pipe, q) (adev)->gfx.funcs->select_me_pipe_q((adev), (me), (pipe), (q))
> +#define amdgpu_gfx_is_gfx_on(adev) (adev)->gfx.funcs->is_gfx_on((adev))
>
> /* Common functions */
> int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> index 2c5e2a4..b8bd194 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> @@ -342,6 +342,18 @@ static int gfx_v9_0_ring_test_ring(struct amdgpu_ring *ring)
> return r;
> }
>
> +static bool gfx_v9_0_is_gfx_on(struct amdgpu_device *adev)
> +{
> + uint32_t reg;
> +
> + reg = RREG32_SOC15(PWR, 0, mmPWR_MISC_CNTL_STATUS);
> + if ((reg & PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS_MASK) ==
> + (0x2 << PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS__SHIFT))
> + return true;
> +
> + return false;
> +}
> +
> static int gfx_v9_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
> {
> struct amdgpu_device *adev = ring->adev;
> @@ -353,6 +365,14 @@ static int gfx_v9_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
> uint32_t tmp;
> long r;
>
> + /*
> + * FIXME: It still has issue to dynamically enable/disable gfxoff at
> + * runtime. So it has to skip the gfx/compute ib test when gfx is
> + * already in "idle" state.
> + */
> + if (!amdgpu_gfx_is_gfx_on(adev))
> + return 0;
> +
> r = amdgpu_device_wb_get(adev, &index);
> if (r) {
> dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r);
> @@ -1085,7 +1105,8 @@ static const struct amdgpu_gfx_funcs gfx_v9_0_gfx_funcs = {
> .read_wave_data = &gfx_v9_0_read_wave_data,
> .read_wave_sgprs = &gfx_v9_0_read_wave_sgprs,
> .read_wave_vgprs = &gfx_v9_0_read_wave_vgprs,
> - .select_me_pipe_q = &gfx_v9_0_select_me_pipe_q
> + .select_me_pipe_q = &gfx_v9_0_select_me_pipe_q,
> + .is_gfx_on = &gfx_v9_0_is_gfx_on
> };
>
> static void gfx_v9_0_gpu_early_init(struct amdgpu_device *adev)
> diff --git a/drivers/gpu/drm/amd/powerplay/hwmgr/smu10_hwmgr.c b/drivers/gpu/drm/amd/powerplay/hwmgr/smu10_hwmgr.c
> index 7712eb6..48c17fb 100644
> --- a/drivers/gpu/drm/amd/powerplay/hwmgr/smu10_hwmgr.c
> +++ b/drivers/gpu/drm/amd/powerplay/hwmgr/smu10_hwmgr.c
> @@ -42,12 +42,6 @@
> #define SMU10_DISPCLK_BYPASS_THRESHOLD 10000 /* 100Mhz */
> #define SMC_RAM_END 0x40000
>
> -#define mmPWR_MISC_CNTL_STATUS 0x0183
> -#define mmPWR_MISC_CNTL_STATUS_BASE_IDX 0
> -#define PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN__SHIFT 0x0
> -#define PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS__SHIFT 0x1
> -#define PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN_MASK 0x00000001L
> -#define PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS_MASK 0x00000006L
>
> static const unsigned long SMU10_Magic = (unsigned long) PHM_Rv_Magic;
>
> @@ -254,28 +248,16 @@ static int smu10_power_off_asic(struct pp_hwmgr *hwmgr)
> return smu10_reset_cc6_data(hwmgr);
> }
>
> -static bool smu10_is_gfx_on(struct pp_hwmgr *hwmgr)
> -{
> - uint32_t reg;
> - struct amdgpu_device *adev = hwmgr->adev;
> -
> - reg = RREG32_SOC15(PWR, 0, mmPWR_MISC_CNTL_STATUS);
> - if ((reg & PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS_MASK) ==
> - (0x2 << PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS__SHIFT))
> - return true;
> -
> - return false;
> -}
> -
> static int smu10_disable_gfx_off(struct pp_hwmgr *hwmgr)
> {
> struct smu10_hwmgr *smu10_data = (struct smu10_hwmgr *)(hwmgr->backend);
> + struct amdgpu_device *adev = hwmgr->adev;
>
> if (smu10_data->gfx_off_controled_by_driver) {
> smum_send_msg_to_smc(hwmgr, PPSMC_MSG_DisableGfxOff);
>
> /* confirm gfx is back to "on" state */
> - while (!smu10_is_gfx_on(hwmgr))
> + while (!amdgpu_gfx_is_gfx_on(adev))
> msleep(1);
> }
>
More information about the amd-gfx
mailing list