drm/amdkfd: New IOCTL to allocate queue GWS (v2)

Add a new kfd ioctl to allocate queue GWS. Queue
GWS is released on queue destroy.

v2: re-introduce this API with the following fixes squashed in:
- drm/amdkfd: fix null pointer dereference on dev
- drm/amdkfd: Return proper error code for gws alloc API
- drm/amdkfd: Remove GPU ID in GWS queue creation
Signed-off-by: NOak Zeng <Oak.Zeng@amd.com>
Reviewed-by: NFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c

@@ -1584,6 +1584,43 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
 	return err;
 }
 
+static int kfd_ioctl_alloc_queue_gws(struct file *filep,
+		struct kfd_process *p, void *data)
+{
+	int retval;
+	struct kfd_ioctl_alloc_queue_gws_args *args = data;
+	struct queue *q;
+	struct kfd_dev *dev;
+
+	if (!hws_gws_support)
+		return -ENODEV;
+
+	mutex_lock(&p->mutex);
+	q = pqm_get_user_queue(&p->pqm, args->queue_id);
+
+	if (q) {
+		dev = q->device;
+	} else {
+		retval = -EINVAL;
+		goto out_unlock;
+	}
+
+	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
+		retval = -ENODEV;
+		goto out_unlock;
+	}
+
+	retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL);
+	mutex_unlock(&p->mutex);
+
+	args->first_gws = 0;
+	return retval;
+
+out_unlock:
+	mutex_unlock(&p->mutex);
+	return retval;
+}
+
 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
 		struct kfd_process *p, void *data)
 {
@@ -1786,6 +1823,8 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
 	AMDKFD_IOCTL_DEF(AMDKFD_IOC_IMPORT_DMABUF,
 				kfd_ioctl_import_dmabuf, 0),
 
+	AMDKFD_IOCTL_DEF(AMDKFD_IOC_ALLOC_QUEUE_GWS,
+			kfd_ioctl_alloc_queue_gws, 0),
 };
 
 #define AMDKFD_CORE_IOCTL_COUNT	ARRAY_SIZE(amdkfd_ioctls)

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

@@ -923,6 +923,8 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
 			void *gws);
 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
 						unsigned int qid);
+struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
+						unsigned int qid);
 int pqm_get_wave_state(struct process_queue_manager *pqm,
 		       unsigned int qid,
 		       void __user *ctl_stack,

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c

@@ -476,6 +476,15 @@ struct kernel_queue *pqm_get_kernel_queue(
 	return NULL;
 }
 
+struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
+					unsigned int qid)
+{
+	struct process_queue_node *pqn;
+
+	pqn = get_queue_by_qid(pqm, qid);
+	return pqn ? pqn->q : NULL;
+}
+
 int pqm_get_wave_state(struct process_queue_manager *pqm,
 		       unsigned int qid,
 		       void __user *ctl_stack,

include/uapi/linux/kfd_ioctl.h

@@ -410,6 +410,20 @@ struct kfd_ioctl_unmap_memory_from_gpu_args {
 	__u32 n_success;		/* to/from KFD */
 };
 
+/* Allocate GWS for specific queue
+ *
+ * @queue_id:    queue's id that GWS is allocated for
+ * @num_gws:     how many GWS to allocate
+ * @first_gws:   index of the first GWS allocated.
+ *               only support contiguous GWS allocation
+ */
+struct kfd_ioctl_alloc_queue_gws_args {
+	__u32 queue_id;		/* to KFD */
+	__u32 num_gws;		/* to KFD */
+	__u32 first_gws;	/* from KFD */
+	__u32 pad;
+};
+
 struct kfd_ioctl_get_dmabuf_info_args {
 	__u64 size;		/* from KFD */
 	__u64 metadata_ptr;	/* to KFD */
@@ -529,7 +543,10 @@ enum kfd_mmio_remap {
 #define AMDKFD_IOC_IMPORT_DMABUF		\
 		AMDKFD_IOWR(0x1D, struct kfd_ioctl_import_dmabuf_args)
 
+#define AMDKFD_IOC_ALLOC_QUEUE_GWS		\
+		AMDKFD_IOWR(0x1E, struct kfd_ioctl_alloc_queue_gws_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1E
+#define AMDKFD_COMMAND_END		0x1F
 
 #endif