drm/amdgpu: dequeue mes scheduler during fini
Subsystems (2)
PATCH v2 1/1 drm/amdgpu: dequeue mes scheduler during fini
Resend to fix coding style issue.
[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.
[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 3 ++ drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 3 ++ drivers/gpu/drm/amd/amdgpu/mes_v11_0.c | 47 +++++++++++++++++++++++-- 3 files changed, 50 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
@@ -130,6 +130,9 @@ struct amdgpu_mes {int (*kiq_hw_init)(struct amdgpu_device *adev);int (*kiq_hw_fini)(struct amdgpu_device *adev); +/* dequeue sched pipe via kiq */ +void (*kiq_dequeue_sched)(struct amdgpu_device *adev); +/* ip specific functions */ const struct amdgpu_mes_funcs *funcs; };
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -0
@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)if (amdgpu_gfx_disable_kcq (adev))DRM_ERROR ("KCQ disable failed\n" ); +if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched) + adev->mes.kiq_dequeue_sched (adev); +amdgpu_mes_kiq_hw_fini (adev); }
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +44 -3
71 lines not on this page · raw message
Signature · 1 line
2.25.1
Christian König
maintainerAm 13.10.22 um 12:39 schrieb YuBiao Wang:
Quoted 112 lines
Resend to fix coding style issue.
[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.Signed-off-by YuBiao Wang <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 3 ++ drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 3 ++ drivers/gpu/drm/amd/amdgpu/mes_v11_0.c | 47 +++++++++++++++++++++++-- 3 files changed, 50 insertions(+), 3 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3
-0@@ -130,6 +130,9 @@ struct amdgpu_mes {int (*kiq_hw_init)(struct amdgpu_device *adev);int (*kiq_hw_fini)(struct amdgpu_device *adev); +/* dequeue sched pipe via kiq */ +void (*kiq_dequeue_sched)(struct amdgpu_device *adev); +/* ip specific functions */ const struct amdgpu_mes_funcs *funcs; };drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3
-0@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)if (amdgpu_gfx_disable_kcq (adev))DRM_ERROR ("KCQ disable failed\n" ); +if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched) + adev->mes.kiq_dequeue_sched (adev); +amdgpu_mes_kiq_hw_fini (adev); }drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +38
-0@@ -44,6 +44,7 @@ MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");static int mes_v11_0_hw_fini (void *handle);static int mes_v11_0_kiq_hw_init (struct amdgpu_device *adev);static int mes_v11_0_kiq_hw_fini (struct amdgpu_device *adev); +static void mes_v11_0_kiq_dequeue_sched (struct amdgpu_device *adev);#define MES_EOP_SIZE 2048@@ -1078,6 +1079,7 @@ static int mes_v11_0_sw_init(void *handle) adev->mes.funcs = &mes_v11_0_funcs; adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init; adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini; + adev->mes.kiq_dequeue_sched = &mes_v11_0_kiq_dequeue_sched; r =amdgpu_mes_init (adev);if (r)@@ -1151,6 +1153,42 @@ static int mes_v11_0_sw_fini(void *handle)return 0 ; } +static void mes_v11_0_kiq_dequeue_sched (struct amdgpu_device *adev) +{ +uint32_t data; +int i; + +mutex_lock (&adev->srbm_mutex); +soc21_grbm_select (adev,3 ,AMDGPU_MES_SCHED_PIPE ,0 ,0 ); + +/* disable the queue if it's active */ +if (RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 ) { +WREG32_SOC15 (GC ,0 , regCP_HQD_DEQUEUE_REQUEST,1 ); +for (i =0 ; i < adev->usec_timeout; i++) { +if (!(RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 )) +break ; +udelay (1 ); + } + } + data =RREG32_SOC15 (GC ,0 , regCP_HQD_PQ_DOORBELL_CONTROL); + data =REG_SET_FIELD (data,CP_HQD_PQ_DOORBELL_CONTROL , +DOORBELL_EN ,0 ); + data =REG_SET_FIELD (data,CP_HQD_PQ_DOORBELL_CONTROL , +DOORBELL_HIT ,1 ); +WREG32_SOC15 (GC ,0 , regCP_HQD_PQ_DOORBELL_CONTROL, data); + +WREG32_SOC15 (GC ,0 , regCP_HQD_PQ_DOORBELL_CONTROL,0 ); + +WREG32_SOC15 (GC ,0 , regCP_HQD_PQ_WPTR_LO,0 ); +WREG32_SOC15 (GC ,0 , regCP_HQD_PQ_WPTR_HI,0 ); +WREG32_SOC15 (GC ,0 , regCP_HQD_PQ_RPTR,0 ); + +soc21_grbm_select (adev,0 ,0 ,0 ,0 ); +mutex_unlock (&adev->srbm_mutex); + + adev->mes.ring.sched.ready = false; +} +static void mes_v11_0_kiq_setting (struct amdgpu_ring *ring) {uint32_t tmp;@@ -1257,9 +1295,12 @@ static int mes_v11_0_hw_init(void *handle)static int mes_v11_0_hw_fini (void *handle)
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c static int mes_v11_0_hw_init(void *handle) · line 1303
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +6
-3{ -struct amdgpu_device *adev = (struct amdgpu_device *)handle; - - adev->mes.ring.sched.ready = false; +/* +* Do not disable mes.ring.sched.ready here, since mes +* is still used to unmap other rings. +* We'll set mes.ring.sched.ready to false when mes ring +* is dequeued. +*/
Maybe drop that comment. The driver should *NEVER* set ring.sched.read to false, but only to true.
The background is that this flag is used if a ring was correctly initialized from the SW side. It's *not* related to the hardware state, but only set to true after the hardware is tested.
We have gotten this wrong so many times and I had to write patches to remove it.
The only exception so far to this is the KIQ ring because here we don't use it for CS. MES might get the same exception, but not as long as we don't need it somewhere.
Regards,
Christian.
Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +0
-0return 0 ; }