gpu: nvgpu: gm20b: support TPC floorsweeping

Bug 1450798

Change-Id: I371537d086ce1088c6d007676c1fe1e2770dd4e3
Signed-off-by: Kevin Huang <kevinh@nvidia.com>
Reviewed-on: http://git-master/r/403877
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>

2 files changed, 71 insertions, 14 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/gk20a_sysfs.c b/drivers/gpu/nvgpu/gk20a/gk20a_sysfs.c
index 7e0183ca..bec18328 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a_sysfs.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a_sysfs.c
@@ -520,6 +520,43 @@ static ssize_t force_idle_read(struct device *device,
 static DEVICE_ATTR(force_idle, ROOTRW, force_idle_read, force_idle_store);
 #endif
 
+static ssize_t tpc_fs_mask_store(struct device *device,
+        struct device_attribute *attr, const char *buf, size_t count)
+{
+        struct platform_device *ndev = to_platform_device(device);
+        struct gk20a *g = get_gk20a(ndev);
+        unsigned long val = 0;
+
+        if (kstrtoul(buf, 10, &val) < 0)
+                return -EINVAL;
+
+        if (val)
+                g->gr.gpc_tpc_mask[0] = val;
+
+        return count;
+}
+
+static ssize_t tpc_fs_mask_read(struct device *device,
+        struct device_attribute *attr, char *buf)
+{
+        struct platform_device *ndev = to_platform_device(device);
+        struct gk20a *g = get_gk20a(ndev);
+        struct gr_gk20a *gr = &g->gr;
+        u32 gpc_index;
+        u32 tpc_fs_mask = 0;
+
+        for (gpc_index = 0; gpc_index < gr->gpc_count; gpc_index++) {
+                if (g->ops.gr.get_gpc_tpc_mask)
+                        tpc_fs_mask |=
+                                g->ops.gr.get_gpc_tpc_mask(g, gpc_index) <<
+                                (gr->max_tpc_per_gpc_count * gpc_index);
+        }
+
+        return sprintf(buf, "0x%x\n", tpc_fs_mask);
+}
+
+static DEVICE_ATTR(tpc_fs_mask, S_IRWXUGO, tpc_fs_mask_read, tpc_fs_mask_store);
+
 void gk20a_remove_sysfs(struct device *dev)
 {
         struct gk20a *g = get_gk20a(to_platform_device(dev));
@@ -540,6 +577,7 @@ void gk20a_remove_sysfs(struct device *dev)
         device_remove_file(dev, &dev_attr_aelpg_param);
         device_remove_file(dev, &dev_attr_aelpg_enable);
         device_remove_file(dev, &dev_attr_allow_all);
+        device_remove_file(dev, &dev_attr_tpc_fs_mask);
 
         if (g->host1x_dev && (dev->parent != &g->host1x_dev->dev))
                 sysfs_remove_link(&dev->kobj, dev_name(dev));
@@ -566,6 +604,7 @@ void gk20a_create_sysfs(struct platform_device *dev)
         error |= device_create_file(&dev->dev, &dev_attr_aelpg_param);
         error |= device_create_file(&dev->dev, &dev_attr_aelpg_enable);
         error |= device_create_file(&dev->dev, &dev_attr_allow_all);
+        error |= device_create_file(&dev->dev, &dev_attr_tpc_fs_mask);
 
         if (g->host1x_dev && (dev->dev.parent != &g->host1x_dev->dev))
                 error |= sysfs_create_link(&g->host1x_dev->dev.kobj,
diff --git a/drivers/gpu/nvgpu/gm20b/gr_gm20b.c b/drivers/gpu/nvgpu/gm20b/gr_gm20b.c
index 25d03736..72500b0e 100644
--- a/drivers/gpu/nvgpu/gm20b/gr_gm20b.c
+++ b/drivers/gpu/nvgpu/gm20b/gr_gm20b.c
@@ -23,6 +23,7 @@
 #include "hw_fb_gm20b.h"
 #include "hw_proj_gm20b.h"
 #include "hw_ctxsw_prog_gm20b.h"
+#include "hw_fuse_gm20b.h"
 
 static void gr_gm20b_init_gpc_mmu(struct gk20a *g)
 {
@@ -478,6 +479,17 @@ static void gr_gm20b_get_sm_dsm_perf_ctrl_regs(struct gk20a *g,
         *ctrl_register_stride = ctxsw_prog_extended_sm_dsm_perf_counter_control_register_stride_v();
 }
 
+static u32 gr_gm20b_get_gpc_tpc_mask(struct gk20a *g, u32 gpc_index)
+{
+        u32 val;
+        struct gr_gk20a *gr = &g->gr;
+
+        /* Toggle the bits of NV_FUSE_STATUS_OPT_TPC_GPC */
+        val = gk20a_readl(g, fuse_status_opt_tpc_gpc_r(gpc_index));
+
+        return (~val) & ((0x1 << gr->max_tpc_per_gpc_count) - 1);
+}
+
 static int gr_gm20b_ctx_state_floorsweep(struct gk20a *g)
 {
         struct gr_gk20a *gr = &g->gr;
@@ -486,6 +498,7 @@ static int gr_gm20b_ctx_state_floorsweep(struct gk20a *g)
         u32 sm_id = 0, gpc_id = 0;
         u32 sm_id_to_gpc_id[proj_scal_max_gpcs_v() * proj_scal_max_tpc_per_gpc_v()];
         u32 tpc_per_gpc;
+        u32 tpc_fs_mask = 0, tpc_sm_id, gpc_tpc_id;
 
         gk20a_dbg_fn("");
 
@@ -556,22 +569,26 @@ static int gr_gm20b_ctx_state_floorsweep(struct gk20a *g)
                      gk20a_readl(g, gr_be0_crop_debug3_r()) |
                      gr_bes_crop_debug3_comp_vdc_4to2_disable_m());
 
-        if (tegra_platform_is_silicon()) {
-                gk20a_writel(g, gr_fe_tpc_fs_r(), gr->pes_tpc_mask[0][0]);
-
-                gk20a_writel(g, gr_cwd_gpc_tpc_id_r(0), gr_cwd_gpc_tpc_id_tpc0_f(0) |
-                                gr_cwd_gpc_tpc_id_tpc1_f(1));
-
-                gk20a_writel(g, gr_cwd_sm_id_r(0), gr_cwd_sm_id_tpc0_f(0) |
-                                gr_cwd_sm_id_tpc1_f(1));
-        } else {
-                gk20a_writel(g, gr_fe_tpc_fs_r(), 1);
-
-                gk20a_writel(g, gr_cwd_gpc_tpc_id_r(0), gr_cwd_gpc_tpc_id_tpc0_f(0));
-
-                gk20a_writel(g, gr_cwd_sm_id_r(0), gr_cwd_sm_id_tpc0_f(0));
+        for (gpc_index = 0; gpc_index < gr->gpc_count; gpc_index++) {
+                tpc_fs_mask |= gr->gpc_tpc_mask[gpc_index] <<
+                                (gr->max_tpc_per_gpc_count * gpc_index);
+        }
+        gk20a_writel(g, gr_fe_tpc_fs_r(), tpc_fs_mask);
 
+        if (tpc_fs_mask & (0x1 << 0)) {
+                tpc_sm_id |= gr_cwd_sm_id_tpc0_f(0);
+                gpc_tpc_id |= gr_cwd_gpc_tpc_id_tpc0_f(0);
+        }
+        if (tpc_fs_mask & (0x1 << 1)) {
+                gpc_tpc_id |= gr_cwd_gpc_tpc_id_tpc1_f(1);
+                tpc_sm_id |= gr_cwd_sm_id_tpc1_f(1);
         }
+        /* Each NV_PGRAPH_PRI_CWD_GPC_TPC_ID can store 4 TPCs.
+         * Since we know TPC number is less than 5. We select
+         * index 0 directly. */
+        gk20a_writel(g, gr_cwd_gpc_tpc_id_r(0), gpc_tpc_id);
+
+        gk20a_writel(g, gr_cwd_sm_id_r(0), tpc_sm_id);
 
         return 0;
 }
@@ -733,4 +750,5 @@ void gm20b_init_gr(struct gpu_ops *gops)
 #else
         gops->gr.load_ctxsw_ucode = gr_gk20a_load_ctxsw_ucode;
 #endif
+        gops->gr.get_gpc_tpc_mask = gr_gm20b_get_gpc_tpc_mask;
 }