drm/msm: Fix a6xx GMU shutdown sequence
authorJordan Crouse <jcrouse@codeaurora.org>
Wed, 5 Feb 2020 20:48:17 +0000 (13:48 -0700)
committerRob Clark <robdclark@chromium.org>
Tue, 11 Feb 2020 15:41:18 +0000 (07:41 -0800)
Commit e812744c5f95 ("drm: msm: a6xx: Add support for A618") missed
updating the VBIF flush in a6xx_gmu_shutdown and instead
inserted the new sequence into a6xx_pm_suspend along with a redundant
GMU idle.

Move a6xx_bus_clear_pending_transactions to a6xx_gmu.c and use it in
the appropriate place in the shutdown routine and remove the redundant
idle call.

v2: Remove newly unused variable that was triggering a warning

Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Reviewed-by: Rob Clark <robdclark@gmail.com>
Fixes: e812744c5f95 ("drm: msm: a6xx: Add support for A618")
Tested-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Rob Clark <robdclark@chromium.org>
drivers/gpu/drm/msm/adreno/a6xx_gmu.c
drivers/gpu/drm/msm/adreno/a6xx_gpu.c

index 983afea..748cd37 100644 (file)
@@ -796,12 +796,41 @@ bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
        return true;
 }
 
+#define GBIF_CLIENT_HALT_MASK             BIT(0)
+#define GBIF_ARB_HALT_MASK                BIT(1)
+
+static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
+{
+       struct msm_gpu *gpu = &adreno_gpu->base;
+
+       if (!a6xx_has_gbif(adreno_gpu)) {
+               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
+               spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
+                                                               0xf) == 0xf);
+               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
+
+               return;
+       }
+
+       /* Halt new client requests on GBIF */
+       gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
+       spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
+                       (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
+
+       /* Halt all AXI requests on GBIF */
+       gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
+       spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
+                       (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
+
+       /* The GBIF halt needs to be explicitly cleared */
+       gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
+}
+
 /* Gracefully try to shut down the GMU and by extension the GPU */
 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
 {
        struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
        struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
-       struct msm_gpu *gpu = &adreno_gpu->base;
        u32 val;
 
        /*
@@ -819,11 +848,7 @@ static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
                        return;
                }
 
-               /* Clear the VBIF pipe before shutting down */
-               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
-               spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) & 0xf)
-                       == 0xf);
-               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
+               a6xx_bus_clear_pending_transactions(adreno_gpu);
 
                /* tell the GMU we want to slumber */
                a6xx_gmu_notify_slumber(gmu);
index 7c449d1..68af241 100644 (file)
@@ -738,39 +738,6 @@ static const u32 a6xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
        REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_A6XX_CP_RB_CNTL),
 };
 
-#define GBIF_CLIENT_HALT_MASK             BIT(0)
-#define GBIF_ARB_HALT_MASK                BIT(1)
-
-static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
-{
-       struct msm_gpu *gpu = &adreno_gpu->base;
-
-       if(!a6xx_has_gbif(adreno_gpu)){
-               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
-               spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
-                                                               0xf) == 0xf);
-               gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
-
-               return;
-       }
-
-       /* Halt new client requests on GBIF */
-       gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
-       spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
-                       (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
-
-       /* Halt all AXI requests on GBIF */
-       gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
-       spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
-                       (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
-
-       /*
-        * GMU needs DDR access in slumber path. Deassert GBIF halt now
-        * to allow for GMU to access system memory.
-        */
-       gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
-}
-
 static int a6xx_pm_resume(struct msm_gpu *gpu)
 {
        struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
@@ -795,16 +762,6 @@ static int a6xx_pm_suspend(struct msm_gpu *gpu)
 
        devfreq_suspend_device(gpu->devfreq.devfreq);
 
-       /*
-        * Make sure the GMU is idle before continuing (because some transitions
-        * may use VBIF
-        */
-       a6xx_gmu_wait_for_idle(&a6xx_gpu->gmu);
-
-       /* Clear the VBIF pipe before shutting down */
-       /* FIXME: This accesses the GPU - do we need to make sure it is on? */
-       a6xx_bus_clear_pending_transactions(adreno_gpu);
-
        return a6xx_gmu_stop(a6xx_gpu);
 }