1 files changed, 60 insertions, 13 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index 6a5986a7..b8753a21 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -60,6 +60,7 @@
 #include "hw_gr_gk20a.h"
 #include "hw_fb_gk20a.h"
 #include "gk20a_scale.h"
+#include "ctxsw_trace_gk20a.h"
 #include "dbg_gpu_gk20a.h"
 #include "gk20a_allocator.h"
 #include "hal.h"
@@ -80,7 +81,7 @@
 /* TODO: Change to e.g. "nvidia-gpu%s" once we have symlinks in place. */
 #define INTERFACE_NAME "nvhost%s-gpu"
 
-#define GK20A_NUM_CDEVS 6
+#define GK20A_NUM_CDEVS 7
 
 #define EMC3D_DEFAULT_RATIO 750
 
@@ -169,6 +170,19 @@ static const struct file_operations gk20a_tsg_ops = {
         .unlocked_ioctl = gk20a_tsg_dev_ioctl,
 };
 
+static const struct file_operations gk20a_ctxsw_ops = {
+        .owner = THIS_MODULE,
+        .release = gk20a_ctxsw_dev_release,
+        .open = gk20a_ctxsw_dev_open,
+#ifdef CONFIG_COMPAT
+        .compat_ioctl = gk20a_ctxsw_dev_ioctl,
+#endif
+        .unlocked_ioctl = gk20a_ctxsw_dev_ioctl,
+        .poll = gk20a_ctxsw_dev_poll,
+        .read = gk20a_ctxsw_dev_read,
+        .mmap = gk20a_ctxsw_dev_mmap,
+};
+
 static inline void sim_writel(struct gk20a *g, u32 r, u32 v)
 {
         writel(v, g->sim.regs+r);
@@ -672,9 +686,6 @@ static int gk20a_init_support(struct platform_device *dev)
         mutex_init(&g->ch_wdt_lock);
         mutex_init(&g->poweroff_lock);
 
-        mutex_init(&g->interleave_lock);
-        g->num_interleaved_channels = 0;
-
         g->remove_support = gk20a_remove_support;
         return 0;
 
@@ -884,6 +895,10 @@ static int gk20a_pm_finalize_poweron(struct device *dev)
                 goto done;
         }
 
+        err = gk20a_ctxsw_trace_init(g);
+        if (err)
+                gk20a_warn(dev, "could not initialize ctxsw tracing");
+
         /* Restore the debug setting */
         g->ops.mm.set_debug_mode(g, g->mmu_debug_ctrl);
 
@@ -1012,6 +1027,11 @@ void gk20a_user_deinit(struct platform_device *dev)
                 cdev_del(&g->tsg.cdev);
         }
 
+        if (g->ctxsw.node) {
+                device_destroy(g->class, g->ctxsw.cdev.dev);
+                cdev_del(&g->ctxsw.cdev);
+        }
+
         if (g->cdev_region)
                 unregister_chrdev_region(g->cdev_region, GK20A_NUM_CDEVS);
 
@@ -1077,6 +1097,15 @@ int gk20a_user_init(struct platform_device *dev)
         if (err)
                 goto fail;
 
+#ifdef CONFIG_GK20A_CTXSW_TRACE
+        err = gk20a_create_device(dev, devno++, "-ctxsw",
+                                  &g->ctxsw.cdev, &g->ctxsw.node,
+                                  &gk20a_ctxsw_ops);
+        if (err)
+                goto fail;
+#endif
+
+
         return 0;
 fail:
         gk20a_user_deinit(dev);
@@ -1400,9 +1429,11 @@ static int gk20a_probe(struct platform_device *dev)
 
         spin_lock_init(&gk20a->mc_enable_lock);
 
+#ifdef CONFIG_RESET_CONTROLLER
         platform->reset_control = devm_reset_control_get(&dev->dev, NULL);
         if (IS_ERR(platform->reset_control))
                 platform->reset_control = NULL;
+#endif
 
         gk20a_debug_init(dev);
 
@@ -1439,14 +1470,11 @@ static int gk20a_probe(struct platform_device *dev)
         if (tegra_platform_is_silicon())
                 gk20a->timeouts_enabled = true;
 
-        gk20a->interleave_high_priority = true;
+        gk20a->runlist_interleave = true;
 
         gk20a->timeslice_low_priority_us = 1300;
         gk20a->timeslice_medium_priority_us = 2600;
-        if (gk20a->interleave_high_priority)
-                gk20a->timeslice_high_priority_us = 3000;
-        else
-                gk20a->timeslice_high_priority_us = 5200;
+        gk20a->timeslice_high_priority_us = 5200;
 
         /* Set up initial power settings. For non-slicon platforms, disable *
          * power features and for silicon platforms, read from platform data */
@@ -1527,16 +1555,17 @@ static int gk20a_probe(struct platform_device *dev)
                                         platform->debugfs,
                                         &gk20a->timeslice_high_priority_us);
 
-        gk20a->debugfs_interleave_high_priority =
-                        debugfs_create_bool("interleave_high_priority",
+        gk20a->debugfs_runlist_interleave =
+                        debugfs_create_bool("runlist_interleave",
                                         S_IRUGO|S_IWUSR,
                                         platform->debugfs,
-                                        &gk20a->interleave_high_priority);
+                                        &gk20a->runlist_interleave);
 
         gr_gk20a_debugfs_init(gk20a);
         gk20a_pmu_debugfs_init(dev);
         gk20a_cde_debugfs_init(dev);
         gk20a_alloc_debugfs_init(dev);
+        gk20a_mm_debugfs_init(dev);
 #endif
 
         gk20a_init_gr(gk20a);
@@ -1558,6 +1587,8 @@ static int __exit gk20a_remove(struct platform_device *dev)
         if (platform->has_cde)
                 gk20a_cde_destroy(g);
 
+        gk20a_ctxsw_trace_cleanup(g);
+
         if (IS_ENABLED(CONFIG_GK20A_DEVFREQ))
                 gk20a_scale_exit(dev);
 
@@ -1774,7 +1805,10 @@ void gk20a_enable(struct gk20a *g, u32 units)
 void gk20a_reset(struct gk20a *g, u32 units)
 {
         gk20a_disable(g, units);
-        udelay(20);
+        if (units & mc_enable_ce2_enabled_f())
+                udelay(500);
+        else
+                udelay(20);
         gk20a_enable(g, units);
 }
 
@@ -2095,6 +2129,19 @@ gk20a_request_firmware(struct gk20a *g, const char *fw_name)
         return fw;
 }
 
+
+u64 gk20a_read_ptimer(struct gk20a *g)
+{
+        u32 time_hi0 = gk20a_readl(g, timer_time_1_r());
+        u32 time_lo = gk20a_readl(g, timer_time_0_r());
+        u32 time_hi1 = gk20a_readl(g, timer_time_1_r());
+        u32 time_hi = (time_lo & (1L << 31)) ? time_hi0 : time_hi1;
+        u64 time = ((u64)time_hi << 32) | time_lo;
+
+        return time;
+}
+
+
 MODULE_LICENSE("GPL v2");
 module_init(gk20a_init);
 module_exit(gk20a_exit);