driver: rknpu: Update rknpu driver, version: 0.9.6

* Add rknpu iommu domain switch support Signed-off-by: Felix Zeng <felix.zeng@rock-chips.com> Change-Id: Ifa756eee01d2d7002ca78a0e1f36ba92a01f4454
radxa · Jun 3, 2024 · 769a605 · 769a605
1 parent 7ae184f
commit 769a605
Show file tree

Hide file tree

Showing 11 changed files with 314 additions and 63 deletions.
diff --git a/drivers/rknpu/include/rknpu_drv.h b/drivers/rknpu/include/rknpu_drv.h
@@ -29,10 +29,10 @@
 
 #define DRIVER_NAME "rknpu"
 #define DRIVER_DESC "RKNPU driver"
-#define DRIVER_DATE "20240226"
+#define DRIVER_DATE "20240322"
 #define DRIVER_MAJOR 0
 #define DRIVER_MINOR 9
-#define DRIVER_PATCHLEVEL 5
+#define DRIVER_PATCHLEVEL 6
 
 #define LOG_TAG "RKNPU"
 
@@ -53,6 +53,8 @@
 #define LOG_DEV_DEBUG(dev, fmt, args...) dev_dbg(dev, LOG_TAG ": " fmt, ##args)
 #define LOG_DEV_ERROR(dev, fmt, args...) dev_err(dev, LOG_TAG ": " fmt, ##args)
 
+#define RKNPU_MAX_IOMMU_DOMAIN_NUM 16
+
 struct rknpu_irqs_data {
 	const char *name;
 	irqreturn_t (*irq_hdl)(int irq, void *ctx);
@@ -120,6 +122,7 @@ struct rknpu_device {
 	spinlock_t irq_lock;
 	struct mutex power_lock;
 	struct mutex reset_lock;
+	struct mutex domain_lock;
 	struct rknpu_subcore_data subcore_datas[RKNPU_MAX_CORES];
 	const struct rknpu_config *config;
 	void __iomem *bw_priority_base;
@@ -163,6 +166,10 @@ struct rknpu_device {
 	void __iomem *nbuf_base_io;
 	struct rknpu_mm *sram_mm;
 	unsigned long power_put_delay;
+	struct iommu_group *iommu_group;
+	int iommu_domain_num;
+	int iommu_domain_id;
+	struct iommu_domain *iommu_domains[RKNPU_MAX_IOMMU_DOMAIN_NUM];
 };
 
 struct rknpu_session {

diff --git a/drivers/rknpu/include/rknpu_gem.h b/drivers/rknpu/include/rknpu_gem.h
@@ -60,6 +60,7 @@ struct rknpu_gem_object {
 	struct page **pages;
 	struct sg_table *sgt;
 	struct drm_mm_node mm_node;
+	int iommu_domain_id;
 };
 
 enum rknpu_cache_type {
@@ -71,7 +72,8 @@ enum rknpu_cache_type {
 struct rknpu_gem_object *rknpu_gem_object_create(struct drm_device *dev,
 						 unsigned int flags,
 						 unsigned long size,
-						 unsigned long sram_size);
+						 unsigned long sram_size,
+						 int iommu_domain_id);
 
 /* destroy a buffer with gem object */
 void rknpu_gem_object_destroy(struct rknpu_gem_object *rknpu_obj);

diff --git a/drivers/rknpu/include/rknpu_ioctl.h b/drivers/rknpu/include/rknpu_ioctl.h
@@ -129,6 +129,8 @@ enum e_rknpu_action {
 	RKNPU_POWER_OFF = 21,
 	RKNPU_GET_TOTAL_SRAM_SIZE = 22,
 	RKNPU_GET_FREE_SRAM_SIZE = 23,
+	RKNPU_GET_IOMMU_DOMAIN_ID = 24,
+	RKNPU_SET_IOMMU_DOMAIN_ID = 25,
 };
 
 /**
@@ -142,6 +144,8 @@ enum e_rknpu_action {
  * @dma_addr: dma address that access by rknpu.
  * @sram_size: user-desired sram memory allocation size.
  *  - this size value would be page-aligned internally.
+ * @iommu_domain_id: iommu domain id
+ * @reserved: just padding to be 64-bit aligned.
  */
 struct rknpu_mem_create {
 	__u32 handle;
@@ -150,6 +154,8 @@ struct rknpu_mem_create {
 	__u64 obj_addr;
 	__u64 dma_addr;
 	__u64 sram_size;
+	__s32 iommu_domain_id;
+	__u32 reserved;
 };
 
 /**
@@ -244,7 +250,8 @@ struct rknpu_subcore_task {
  * @task_counter: task counter
  * @priority: submit priority
  * @task_obj_addr: address of task object
- * @regcfg_obj_addr: address of register config object
+ * @iommu_domain_id: iommu domain id
+ * @reserved: just padding to be 64-bit aligned.
  * @task_base_addr: task base address
  * @hw_elapse_time: hardware elapse time
  * @core_mask: core mask of rknpu
@@ -260,7 +267,8 @@ struct rknpu_submit {
 	__u32 task_counter;
 	__s32 priority;
 	__u64 task_obj_addr;
-	__u64 regcfg_obj_addr;
+	__u32 iommu_domain_id;
+	__u32 reserved;
 	__u64 task_base_addr;
 	__s64 hw_elapse_time;
 	__u32 core_mask;

diff --git a/drivers/rknpu/include/rknpu_iommu.h b/drivers/rknpu/include/rknpu_iommu.h
@@ -37,4 +37,12 @@ dma_addr_t rknpu_iommu_dma_alloc_iova(struct iommu_domain *domain, size_t size,
 void rknpu_iommu_dma_free_iova(struct rknpu_iommu_dma_cookie *cookie,
 			       dma_addr_t iova, size_t size);
 
+int rknpu_iommu_init_domain(struct rknpu_device *rknpu_dev);
+int rknpu_iommu_switch_domain(struct rknpu_device *rknpu_dev, int domain_id);
+void rknpu_iommu_free_domains(struct rknpu_device *rknpu_dev);
+
+#if KERNEL_VERSION(5, 10, 0) < LINUX_VERSION_CODE
+int iommu_get_dma_cookie(struct iommu_domain *domain);
+#endif
+
 #endif
diff --git a/drivers/rknpu/include/rknpu_job.h b/drivers/rknpu/include/rknpu_job.h
@@ -48,6 +48,7 @@ struct rknpu_job {
 	ktime_t hw_recoder_time;
 	ktime_t hw_elapse_time;
 	atomic_t submit_count[RKNPU_MAX_CORES];
+	int iommu_domain_id;
 };
 
 irqreturn_t rknpu_core0_irq_handler(int irq, void *data);

diff --git a/drivers/rknpu/include/rknpu_mem.h b/drivers/rknpu/include/rknpu_mem.h
@@ -37,10 +37,10 @@ struct rknpu_mem_object {
 	unsigned int owner;
 };
 
-int rknpu_mem_create_ioctl(struct rknpu_device *rknpu_dev, unsigned long data,
-			   struct file *file);
-int rknpu_mem_destroy_ioctl(struct rknpu_device *rknpu_dev, unsigned long data,
-			    struct file *file);
+int rknpu_mem_create_ioctl(struct rknpu_device *rknpu_dev, struct file *file,
+			   unsigned int cmd, unsigned long data);
+int rknpu_mem_destroy_ioctl(struct rknpu_device *rknpu_dev, struct file *file,
+			    unsigned long data);
 int rknpu_mem_sync_ioctl(struct rknpu_device *rknpu_dev, unsigned long data);
 
 #endif
diff --git a/drivers/rknpu/rknpu_drv.c b/drivers/rknpu/rknpu_drv.c
@@ -43,6 +43,7 @@
 #include "rknpu_drv.h"
 #include "rknpu_gem.h"
 #include "rknpu_devfreq.h"
+#include "rknpu_iommu.h"
 
 #ifdef CONFIG_ROCKCHIP_RKNPU_DRM_GEM
 #include <drm/drm_device.h>
@@ -375,6 +376,15 @@ static int rknpu_action(struct rknpu_device *rknpu_dev,
 			args->value = 0;
 		ret = 0;
 		break;
+	case RKNPU_GET_IOMMU_DOMAIN_ID:
+		args->value = rknpu_dev->iommu_domain_id;
+		ret = 0;
+		break;
+	case RKNPU_SET_IOMMU_DOMAIN_ID: {
+		ret = rknpu_iommu_switch_domain(rknpu_dev,
+						*(int32_t *)&args->value);
+		break;
+	}
 	default:
 		ret = -EINVAL;
 		break;
@@ -464,7 +474,7 @@ static int rknpu_action_ioctl(struct rknpu_device *rknpu_dev,
 	return ret;
 }
 
-static long rknpu_ioctl(struct file *file, uint32_t cmd, unsigned long arg)
+static long rknpu_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 {
 	long ret = -EINVAL;
 	struct rknpu_device *rknpu_dev = NULL;
@@ -476,22 +486,22 @@ static long rknpu_ioctl(struct file *file, uint32_t cmd, unsigned long arg)
 
 	rknpu_power_get(rknpu_dev);
 
-	switch (cmd) {
-	case IOCTL_RKNPU_ACTION:
+	switch (_IOC_NR(cmd)) {
+	case RKNPU_ACTION:
 		ret = rknpu_action_ioctl(rknpu_dev, arg);
 		break;
-	case IOCTL_RKNPU_SUBMIT:
+	case RKNPU_SUBMIT:
 		ret = rknpu_submit_ioctl(rknpu_dev, arg);
 		break;
-	case IOCTL_RKNPU_MEM_CREATE:
-		ret = rknpu_mem_create_ioctl(rknpu_dev, arg, file);
+	case RKNPU_MEM_CREATE:
+		ret = rknpu_mem_create_ioctl(rknpu_dev, file, cmd, arg);
 		break;
 	case RKNPU_MEM_MAP:
 		break;
-	case IOCTL_RKNPU_MEM_DESTROY:
-		ret = rknpu_mem_destroy_ioctl(rknpu_dev, arg, file);
+	case RKNPU_MEM_DESTROY:
+		ret = rknpu_mem_destroy_ioctl(rknpu_dev, file, arg);
 		break;
-	case IOCTL_RKNPU_MEM_SYNC:
+	case RKNPU_MEM_SYNC:
 		ret = rknpu_mem_sync_ioctl(rknpu_dev, arg);
 		break;
 	default:
@@ -1171,7 +1181,11 @@ static int rknpu_probe(struct platform_device *pdev)
 	rknpu_dev->dev = dev;
 
 	rknpu_dev->iommu_en = rknpu_is_iommu_enable(dev);
-	if (!rknpu_dev->iommu_en) {
+	if (rknpu_dev->iommu_en) {
+		rknpu_dev->iommu_group = iommu_group_get(dev);
+		if (!rknpu_dev->iommu_group)
+			return -EINVAL;
+	} else {
 		/* Initialize reserved memory resources */
 		ret = of_reserved_mem_device_init(dev);
 		if (!ret) {
@@ -1226,6 +1240,7 @@ static int rknpu_probe(struct platform_device *pdev)
 	spin_lock_init(&rknpu_dev->irq_lock);
 	mutex_init(&rknpu_dev->power_lock);
 	mutex_init(&rknpu_dev->reset_lock);
+	mutex_init(&rknpu_dev->domain_lock);
 	for (i = 0; i < config->num_irqs; i++) {
 		INIT_LIST_HEAD(&rknpu_dev->subcore_datas[i].todo_list);
 		init_waitqueue_head(&rknpu_dev->subcore_datas[i].job_done_wq);
@@ -1365,6 +1380,9 @@ static int rknpu_probe(struct platform_device *pdev)
 	    rknpu_dev->config->nbuf_size > 0)
 		rknpu_find_nbuf_resource(rknpu_dev);
 
+	if (rknpu_dev->iommu_en)
+		rknpu_iommu_init_domain(rknpu_dev);
+
 	rknpu_power_off(rknpu_dev);
 	atomic_set(&rknpu_dev->power_refcount, 0);
 	atomic_set(&rknpu_dev->cmdline_power_refcount, 0);
@@ -1401,9 +1419,6 @@ static int rknpu_remove(struct platform_device *pdev)
 	cancel_delayed_work_sync(&rknpu_dev->power_off_work);
 	destroy_workqueue(rknpu_dev->power_off_wq);
 
-	if (IS_ENABLED(CONFIG_ROCKCHIP_RKNPU_SRAM) && rknpu_dev->sram_mm)
-		rknpu_mm_destroy(rknpu_dev->sram_mm);
-
 	rknpu_debugger_remove(rknpu_dev);
 	rknpu_cancel_timer(rknpu_dev);
 
@@ -1412,6 +1427,14 @@ static int rknpu_remove(struct platform_device *pdev)
 		WARN_ON(!list_empty(&rknpu_dev->subcore_datas[i].todo_list));
 	}
 
+	if (IS_ENABLED(CONFIG_ROCKCHIP_RKNPU_SRAM) && rknpu_dev->sram_mm)
+		rknpu_mm_destroy(rknpu_dev->sram_mm);
+
+	if (rknpu_dev->iommu_en) {
+		rknpu_iommu_free_domains(rknpu_dev);
+		iommu_group_put(rknpu_dev->iommu_group);
+	}
+
 #ifdef CONFIG_ROCKCHIP_RKNPU_DRM_GEM
 	rknpu_drm_remove(rknpu_dev);
 #endif