10 files changed, 91 insertions, 61 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/ce2_gk20a.c b/drivers/gpu/nvgpu/gk20a/ce2_gk20a.c
index 921ee6f8..5dfd2309 100644
--- a/drivers/gpu/nvgpu/gk20a/ce2_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/ce2_gk20a.c
@@ -690,6 +690,12 @@ void gk20a_ce_delete_context(struct device *dev,
                 u32 ce_ctx_id)
 {
         struct gk20a *g = gk20a_from_dev(dev);
+        gk20a_ce_delete_context_priv(g, ce_ctx_id);
+}
+
+void gk20a_ce_delete_context_priv(struct gk20a *g,
+                u32 ce_ctx_id)
+{
         struct gk20a_ce_app *ce_app = &g->ce_app;
         struct gk20a_gpu_ctx *ce_ctx, *ce_ctx_save;
 
diff --git a/drivers/gpu/nvgpu/gk20a/ce2_gk20a.h b/drivers/gpu/nvgpu/gk20a/ce2_gk20a.h
index 5cdd233e..7ecf130f 100644
--- a/drivers/gpu/nvgpu/gk20a/ce2_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/ce2_gk20a.h
@@ -144,9 +144,12 @@ int gk20a_ce_execute_ops(struct device *dev,
                 struct gk20a_fence *gk20a_fence_in,
                 u32 submit_flags,
                 struct gk20a_fence **gk20a_fence_out);
+void gk20a_ce_delete_context_priv(struct gk20a *g,
+                u32 ce_ctx_id);
 void gk20a_ce_delete_context(struct device *dev,
                 u32 ce_ctx_id);
 
+
 #ifdef CONFIG_DEBUG_FS
 /* CE app debugfs api */
 void gk20a_ce_debugfs_init(struct device *dev);
diff --git a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
index 0249a1c6..d0d38f83 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
@@ -122,7 +122,8 @@ static struct channel_gk20a *allocate_channel(struct fifo_gk20a *f)
 static void free_channel(struct fifo_gk20a *f,
                 struct channel_gk20a *ch)
 {
-        struct gk20a_platform *platform = gk20a_get_platform(f->g->dev);
+        struct gk20a_platform *platform;
+        struct gk20a *g = f->g;
 
         trace_gk20a_release_used_channel(ch->hw_chid);
         /* refcount is zero here and channel is in a freed/dead state */
@@ -132,10 +133,18 @@ static void free_channel(struct fifo_gk20a *f,
         f->used_channels--;
         nvgpu_mutex_release(&f->free_chs_mutex);
 
-        if (platform->aggressive_sync_destroy_thresh &&
+        /*
+         * On teardown it is not possible to dereference platform, but ignoring
+         * this is fine then because no new channels would be created.
+         */
+        if (!g->driver_is_dying) {
+                platform = gk20a_get_platform(g->dev);
+
+                if (platform->aggressive_sync_destroy_thresh &&
                         (f->used_channels <
                          platform->aggressive_sync_destroy_thresh))
-                platform->aggressive_sync_destroy = false;
+                        platform->aggressive_sync_destroy = false;
+        }
 }
 
 int channel_gk20a_commit_va(struct channel_gk20a *c)
@@ -3016,7 +3025,12 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c,
         bool need_deferred_cleanup = false;
         struct nvgpu_gpfifo __user *user_gpfifo = args ?
                 (struct nvgpu_gpfifo __user *)(uintptr_t)args->gpfifo : NULL;
-        struct gk20a_platform *platform = gk20a_get_platform(d);
+        struct gk20a_platform *platform;
+
+        if (g->driver_is_dying)
+                return -ENODEV;
+
+        platform = gk20a_get_platform(d);
 
         if (c->has_timedout)
                 return -ETIMEDOUT;
diff --git a/drivers/gpu/nvgpu/gk20a/ctrl_gk20a.c b/drivers/gpu/nvgpu/gk20a/ctrl_gk20a.c
index c87226c8..1ae42337 100644
--- a/drivers/gpu/nvgpu/gk20a/ctrl_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/ctrl_gk20a.c
@@ -1344,8 +1344,7 @@ static int nvgpu_gpu_set_therm_alert_limit(struct gk20a *g,
 long gk20a_ctrl_dev_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 {
         struct gk20a_ctrl_priv *priv = filp->private_data;
-        struct device *dev = priv->dev;
-        struct gk20a *g = get_gk20a(dev);
+        struct gk20a *g = priv->g;
         struct nvgpu_gpu_zcull_get_ctx_size_args *get_ctx_size_args;
         struct nvgpu_gpu_zcull_get_info_args *get_info_args;
         struct nvgpu_gpu_zbc_set_table_args *set_table_args;
diff --git a/drivers/gpu/nvgpu/gk20a/dbg_gpu_gk20a.c b/drivers/gpu/nvgpu/gk20a/dbg_gpu_gk20a.c
index db534318..165bcf46 100644
--- a/drivers/gpu/nvgpu/gk20a/dbg_gpu_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/dbg_gpu_gk20a.c
@@ -665,7 +665,7 @@ static int nvgpu_dbg_gpu_ioctl_timeout(struct dbg_session_gk20a *dbg_s,
                          struct nvgpu_dbg_gpu_timeout_args *args)
 {
         int err;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
 
         gk20a_dbg_fn("powergate mode = %d", args->enable);
 
@@ -680,7 +680,7 @@ static void nvgpu_dbg_gpu_ioctl_get_timeout(struct dbg_session_gk20a *dbg_s,
                          struct nvgpu_dbg_gpu_timeout_args *args)
 {
         int status;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
 
         nvgpu_mutex_acquire(&g->dbg_sessions_lock);
         status = g->timeouts_enabled;
@@ -711,7 +711,7 @@ static int nvgpu_dbg_gpu_ioctl_read_single_sm_error_state(
                 struct dbg_session_gk20a *dbg_s,
                 struct nvgpu_dbg_gpu_read_single_sm_error_state_args *args)
 {
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct gr_gk20a *gr = &g->gr;
         struct nvgpu_dbg_gpu_sm_error_state_record *sm_error_state;
         u32 sm_id;
@@ -750,7 +750,7 @@ static int nvgpu_dbg_gpu_ioctl_clear_single_sm_error_state(
                 struct dbg_session_gk20a *dbg_s,
                 struct nvgpu_dbg_gpu_clear_single_sm_error_state_args *args)
 {
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct gr_gk20a *gr = &g->gr;
         u32 sm_id;
         struct channel_gk20a *ch;
@@ -781,7 +781,7 @@ static int nvgpu_dbg_gpu_ioctl_write_single_sm_error_state(
                 struct dbg_session_gk20a *dbg_s,
                 struct nvgpu_dbg_gpu_write_single_sm_error_state_args *args)
 {
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct gr_gk20a *gr = &g->gr;
         u32 sm_id;
         struct channel_gk20a *ch;
@@ -952,7 +952,7 @@ long gk20a_dbg_gpu_dev_ioctl(struct file *filp, unsigned int cmd,
                              unsigned long arg)
 {
         struct dbg_session_gk20a *dbg_s = filp->private_data;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         u8 buf[NVGPU_DBG_GPU_IOCTL_MAX_ARG_SIZE];
         int err = 0;
 
@@ -1141,7 +1141,7 @@ static int nvgpu_ioctl_channel_reg_ops(struct dbg_session_gk20a *dbg_s,
         bool is_pg_disabled = false;
 
         struct device *dev = dbg_s->dev;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct channel_gk20a *ch;
 
         gk20a_dbg_fn("%d ops, max fragment %d", args->num_ops, g->dbg_regops_tmp_buf_ops);
@@ -1257,7 +1257,7 @@ static int nvgpu_ioctl_channel_reg_ops(struct dbg_session_gk20a *dbg_s,
 static int dbg_set_powergate(struct dbg_session_gk20a *dbg_s, u32  powermode)
 {
         int err = 0;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
 
          /* This function must be called with g->dbg_sessions_lock held */
 
@@ -1360,7 +1360,7 @@ static int nvgpu_ioctl_powergate_gk20a(struct dbg_session_gk20a *dbg_s,
                                 struct nvgpu_dbg_gpu_powergate_args *args)
 {
         int err;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         gk20a_dbg_fn("%s  powergate mode = %d",
                       dev_name(dbg_s->dev), args->mode);
 
@@ -1374,7 +1374,7 @@ static int nvgpu_dbg_gpu_ioctl_smpc_ctxsw_mode(struct dbg_session_gk20a *dbg_s,
                                struct nvgpu_dbg_gpu_smpc_ctxsw_mode_args *args)
 {
         int err;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct channel_gk20a *ch_gk20a;
 
         gk20a_dbg_fn("%s smpc ctxsw mode = %d",
@@ -1416,7 +1416,7 @@ static int nvgpu_dbg_gpu_ioctl_hwpm_ctxsw_mode(struct dbg_session_gk20a *dbg_s,
                                struct nvgpu_dbg_gpu_hwpm_ctxsw_mode_args *args)
 {
         int err;
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct channel_gk20a *ch_gk20a;
 
         gk20a_dbg_fn("%s pm ctxsw mode = %d",
@@ -1468,7 +1468,7 @@ static int nvgpu_dbg_gpu_ioctl_suspend_resume_sm(
                 struct dbg_session_gk20a *dbg_s,
                 struct nvgpu_dbg_gpu_suspend_resume_all_sms_args *args)
 {
-        struct gk20a *g = get_gk20a(dbg_s->dev);
+        struct gk20a *g = dbg_s->g;
         struct channel_gk20a *ch;
         int err = 0, action = args->mode;
 
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index 2046c08e..694f0e93 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -334,26 +334,14 @@ static irqreturn_t gk20a_intr_thread_stall(int irq, void *dev_id)
         return g->ops.mc.isr_thread_stall(g);
 }
 
-void gk20a_remove_support(struct device *dev)
+void gk20a_remove_support(struct gk20a *g)
 {
-        struct gk20a *g = get_gk20a(dev);
-
 #ifdef CONFIG_TEGRA_COMMON
         tegra_unregister_idle_unidle();
 #endif
         if (g->dbg_regops_tmp_buf)
                 kfree(g->dbg_regops_tmp_buf);
 
-        nvgpu_wait_for_deferred_interrupts(g);
-
-        gk20a_channel_cancel_pending_sema_waits(g);
-
-        if (g->nonstall_work_queue) {
-                cancel_work_sync(&g->nonstall_fn_work);
-                destroy_workqueue(g->nonstall_work_queue);
-                g->nonstall_work_queue = NULL;
-        }
-
         if (g->pmu.remove_support)
                 g->pmu.remove_support(&g->pmu);
 
@@ -1251,6 +1239,11 @@ static int gk20a_probe(struct platform_device *dev)
         if (gk20a->irq_stall != gk20a->irq_nonstall)
                 disable_irq(gk20a->irq_nonstall);
 
+        /*
+         * is_fmodel needs to be in gk20a struct for deferred teardown
+         */
+        gk20a->is_fmodel = platform->is_fmodel;
+
         err = gk20a_init_support(dev);
         if (err)
                 return err;
@@ -1297,11 +1290,6 @@ static int __exit gk20a_remove(struct platform_device *pdev)
         if (IS_ENABLED(CONFIG_GK20A_DEVFREQ))
                 gk20a_scale_exit(dev);
 
-        if (g->remove_support)
-                g->remove_support(dev);
-
-        gk20a_ce_destroy(g);
-
 #ifdef CONFIG_ARCH_TEGRA_18x_SOC
         nvgpu_clk_arb_cleanup_arbiter(g);
 #endif
@@ -1390,7 +1378,21 @@ void gk20a_busy_noresume(struct device *dev)
 void gk20a_driver_start_unload(struct gk20a *g)
 {
         gk20a_dbg(gpu_dbg_shutdown, "Driver is now going down!\n");
+
+        down_write(&g->busy_lock);
         g->driver_is_dying = 1;
+        up_write(&g->busy_lock);
+
+        gk20a_wait_for_idle(g->dev);
+
+        nvgpu_wait_for_deferred_interrupts(g);
+        gk20a_channel_cancel_pending_sema_waits(g);
+
+        if (g->nonstall_work_queue) {
+                cancel_work_sync(&g->nonstall_fn_work);
+                destroy_workqueue(g->nonstall_work_queue);
+                g->nonstall_work_queue = NULL;
+        }
 }
 
 int gk20a_wait_for_idle(struct device *dev)
@@ -1859,6 +1861,12 @@ static void gk20a_free_cb(struct kref *refcount)
                 struct gk20a, refcount);
 
         gk20a_dbg(gpu_dbg_shutdown, "Freeing GK20A struct!");
+
+        gk20a_ce_destroy(g);
+
+        if (g->remove_support)
+                g->remove_support(g);
+
         kfree(g);
 }
 
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.h b/drivers/gpu/nvgpu/gk20a/gk20a.h
index 30f1b371..f4ca5649 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.h
@@ -877,6 +877,9 @@ struct gk20a {
         atomic_t nonstall_ops;
         struct work_struct nonstall_fn_work;
         struct workqueue_struct *nonstall_work_queue;
+
+        bool is_fmodel;
+
         struct kref refcount;
 
         struct resource *reg_mem;
@@ -987,7 +990,7 @@ struct gk20a {
         bool global_profiler_reservation_held;
         int profiler_reservation_count;
 
-        void (*remove_support)(struct device *);
+        void (*remove_support)(struct gk20a *);
 
         u64 pg_ingating_time_us;
         u64 pg_ungating_time_us;
@@ -1455,7 +1458,7 @@ extern struct class nvgpu_class;
 
 int gk20a_pm_init(struct device *dev);
 int gk20a_pm_finalize_poweron(struct device *dev);
-void gk20a_remove_support(struct device *dev);
+void gk20a_remove_support(struct gk20a *g);
 
 static inline struct tsg_gk20a *tsg_gk20a_from_ch(struct channel_gk20a *ch)
 {
diff --git a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
index 7b08387e..9e6dc74c 100644
--- a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
@@ -809,7 +809,7 @@ static void gk20a_remove_mm_ce_support(struct mm_gk20a *mm)
         struct gk20a *g = gk20a_from_mm(mm);
 
         if (mm->vidmem.ce_ctx_id != (u32)~0)
-                gk20a_ce_delete_context(g->dev, mm->vidmem.ce_ctx_id);
+                gk20a_ce_delete_context_priv(g, mm->vidmem.ce_ctx_id);
 
         mm->vidmem.ce_ctx_id = (u32)~0;
 
@@ -1220,11 +1220,10 @@ static int alloc_gmmu_pages(struct vm_gk20a *vm, u32 order,
         u32 num_pages = 1 << order;
         u32 len = num_pages * PAGE_SIZE;
         int err;
-        struct gk20a_platform *platform = dev_get_drvdata(g->dev);
 
         gk20a_dbg_fn("");
 
-        if (platform->is_fmodel)
+        if (g->is_fmodel)
                 return alloc_gmmu_phys_pages(vm, order, entry);
 
         /*
@@ -1250,7 +1249,6 @@ void free_gmmu_pages(struct vm_gk20a *vm,
                      struct gk20a_mm_entry *entry)
 {
         struct gk20a *g = gk20a_from_vm(vm);
-        struct gk20a_platform *platform = dev_get_drvdata(g->dev);
 
         gk20a_dbg_fn("");
 
@@ -1260,7 +1258,7 @@ void free_gmmu_pages(struct vm_gk20a *vm,
         if (entry->woffset) /* fake shadow mem */
                 return;
 
-        if (platform->is_fmodel) {
+        if (g->is_fmodel) {
                 free_gmmu_phys_pages(vm, entry);
                 return;
         }
@@ -1270,11 +1268,9 @@ void free_gmmu_pages(struct vm_gk20a *vm,
 
 int map_gmmu_pages(struct gk20a *g, struct gk20a_mm_entry *entry)
 {
-        struct gk20a_platform *platform = dev_get_drvdata(g->dev);
-
         gk20a_dbg_fn("");
 
-        if (platform->is_fmodel)
+        if (g->is_fmodel)
                 return map_gmmu_phys_pages(entry);
 
         if (IS_ENABLED(CONFIG_ARM64)) {
@@ -1296,11 +1292,9 @@ int map_gmmu_pages(struct gk20a *g, struct gk20a_mm_entry *entry)
 
 void unmap_gmmu_pages(struct gk20a *g, struct gk20a_mm_entry *entry)
 {
-        struct gk20a_platform *platform = dev_get_drvdata(g->dev);
-
         gk20a_dbg_fn("");
 
-        if (platform->is_fmodel) {
+        if (g->is_fmodel) {
                 unmap_gmmu_phys_pages(entry);
                 return;
         }
@@ -4070,6 +4064,7 @@ static void gk20a_vm_remove_support_nofree(struct vm_gk20a *vm)
         struct mapped_buffer_node *mapped_buffer;
         struct vm_reserved_va_node *va_node, *va_node_tmp;
         struct rb_node *node;
+        struct gk20a *g = vm->mm->g;
 
         gk20a_dbg_fn("");
 
@@ -4078,7 +4073,7 @@ static void gk20a_vm_remove_support_nofree(struct vm_gk20a *vm)
          * pool involves unmapping a GMMU mapping which means aquiring the
          * update_gmmu_lock.
          */
-        if (!gk20a_platform_has_syncpoints(gk20a_from_vm(vm)->dev)) {
+        if (!(g->gpu_characteristics.flags & NVGPU_GPU_FLAGS_HAS_SYNCPOINTS)) {
                 if (vm->sema_pool) {
                         nvgpu_semaphore_pool_unmap(vm->sema_pool, vm);
                         nvgpu_semaphore_pool_put(vm->sema_pool);
@@ -4172,7 +4167,7 @@ static int gk20a_init_sema_pool(struct vm_gk20a *vm)
         /*
          * Don't waste the memory on semaphores if we don't need them.
          */
-        if (gk20a_platform_has_syncpoints(g->dev))
+        if (g->gpu_characteristics.flags & NVGPU_GPU_FLAGS_HAS_SYNCPOINTS)
                 return 0;
 
         if (vm->sema_pool)
diff --git a/drivers/gpu/nvgpu/pci.c b/drivers/gpu/nvgpu/pci.c
index 114e9af7..a7899f7e 100644
--- a/drivers/gpu/nvgpu/pci.c
+++ b/drivers/gpu/nvgpu/pci.c
@@ -393,6 +393,11 @@ static int nvgpu_pci_probe(struct pci_dev *pdev,
         }
         disable_irq(g->irq_stall);
 
+        /*
+         * is_fmodel needs to be in gk20a struct for deferred teardown
+         */
+        g->is_fmodel = platform->is_fmodel;
+
         err = nvgpu_pci_init_support(pdev);
         if (err)
                 return err;
@@ -426,7 +431,6 @@ static void nvgpu_pci_remove(struct pci_dev *pdev)
         struct gk20a *g = get_gk20a(&pdev->dev);
 
         gk20a_dbg(gpu_dbg_shutdown, "Removing nvgpu driver!\n");
-        gk20a_driver_start_unload(g);
 
         if (g->irqs_enabled)
                 disable_irq(g->irq_stall);
@@ -445,7 +449,7 @@ static void nvgpu_pci_remove(struct pci_dev *pdev)
          * Wait for the driver to finish up all the IOCTLs it's working on
          * before cleaning up the driver's data structures.
          */
-        gk20a_wait_for_idle(&pdev->dev);
+        gk20a_driver_start_unload(g);
         gk20a_dbg(gpu_dbg_shutdown, "Driver idle.\n");
 
 #ifdef CONFIG_ARCH_TEGRA_18x_SOC
@@ -455,9 +459,6 @@ static void nvgpu_pci_remove(struct pci_dev *pdev)
         gk20a_user_deinit(g->dev, &nvgpu_pci_class);
         gk20a_dbg(gpu_dbg_shutdown, "User de-init done.\b");
 
-        if (g->remove_support)
-                g->remove_support(g->dev);
-
         debugfs_remove_recursive(platform->debugfs);
         debugfs_remove_recursive(platform->debugfs_alias);
 
diff --git a/drivers/gpu/nvgpu/vgpu/vgpu.c b/drivers/gpu/nvgpu/vgpu/vgpu.c
index 72606952..df793be7 100644
--- a/drivers/gpu/nvgpu/vgpu/vgpu.c
+++ b/drivers/gpu/nvgpu/vgpu/vgpu.c
@@ -191,10 +191,9 @@ static int vgpu_intr_thread(void *dev_id)
         return 0;
 }
 
-static void vgpu_remove_support(struct device *dev)
+static void vgpu_remove_support(struct gk20a *g)
 {
-        struct gk20a *g = get_gk20a(dev);
-        struct vgpu_priv_data *priv = vgpu_get_priv_data_from_dev(dev);
+        struct vgpu_priv_data *priv = vgpu_get_priv_data_from_dev(g->dev);
         struct tegra_vgpu_intr_msg msg;
         int err;
 
@@ -265,7 +264,7 @@ static int vgpu_init_support(struct platform_device *pdev)
         return 0;
 
  fail:
-        vgpu_remove_support(&pdev->dev);
+        vgpu_remove_support(g);
         return err;
 }
 
@@ -571,6 +570,8 @@ int vgpu_probe(struct platform_device *pdev)
         platform->vgpu_priv = priv;
         gk20a->dev = dev;
 
+        gk20a->is_fmodel = platform->is_fmodel;
+
         nvgpu_kmem_init(gk20a);
 
         err = gk20a_user_init(dev, INTERFACE_NAME, &nvgpu_class);
@@ -653,7 +654,7 @@ int vgpu_remove(struct platform_device *pdev)
 
         vgpu_pm_qos_remove(dev);
         if (g->remove_support)
-                g->remove_support(dev);
+                g->remove_support(g);
 
         vgpu_comm_deinit();
         gk20a_sched_ctrl_cleanup(g);