[PATCH 5/8] drm/amdkfd: Add function to set queue gws

[PATCH 5/8] drm/amdkfd: Add function to set queue gws

[Zeng, Oak]

Add functions in process queue manager to
set/get queue gws. Also set process's number
of gws used. Currently only one queue in
process can use and use all gws.

Change-Id: I03e480c8692db3eabfc3a188cce8904d5d962ab7
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  6 +++
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 50 ++++++++++++++++++++++
 2 files changed, 56 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 5969e37..40a5c67 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -454,6 +454,9 @@ struct queue_properties {
  *
  * @device: The kfd device that created this queue.
  *
+ * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
+ * otherwise.
+ *
  * This structure represents user mode compute queues.
  * It contains all the necessary data to handle such queues.
  *
@@ -475,6 +478,7 @@ struct queue {
 
 	struct kfd_process	*process;
 	struct kfd_dev		*device;
+	void *gws;
 };
 
 /*
@@ -868,6 +872,8 @@ int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
 			struct queue_properties *p);
 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
 			struct queue_properties *p);
+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);
 int pqm_get_wave_state(struct process_queue_manager *pqm,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
index e652e25..8e45296 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
@@ -26,6 +26,7 @@
 #include "kfd_device_queue_manager.h"
 #include "kfd_priv.h"
 #include "kfd_kernel_queue.h"
+#include "amdgpu_amdkfd.h"
 
 static inline struct process_queue_node *get_queue_by_qid(
 			struct process_queue_manager *pqm, unsigned int qid)
@@ -74,6 +75,55 @@ void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
 	pdd->already_dequeued = true;
 }
 
+int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
+			void *gws)
+{
+	struct kfd_dev *dev = NULL;
+	struct process_queue_node *pqn;
+	struct kfd_process_device *pdd;
+
+	pqn = get_queue_by_qid(pqm, qid);
+	if (!pqn) {
+		pr_err("Queue id does not match any known queue\n");
+		return -EINVAL;
+	}
+
+	if (pqn->q)
+		dev = pqn->q->device;
+	if (WARN_ON(!dev))
+		return -ENODEV;
+
+	pdd = kfd_get_process_device_data(dev, pqm->process);
+	if (!pdd) {
+		pr_err("Process device data doesn't exist\n");
+		return -EINVAL;
+	}
+
+	/* Only allow one queue per process can have GWS assigned */
+	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
+		if (pqn->q && pqn->q->gws)
+			return -EINVAL;
+	}
+
+	pqn->q->gws = gws;
+	pdd->qpd.num_gws = gws ? amdgpu_amdkfd_get_num_gws(dev->kgd) : 0;
+	return 0;
+}
+
+static void *pqm_get_gws(struct process_queue_manager *pqm, unsigned int qid)
+{
+	struct process_queue_node *pqn;
+
+	pqn = get_queue_by_qid(pqm, qid);
+	if (!pqn) {
+		pr_debug("No queue %d exists for get gws operation\n", qid);
+		return NULL;
+	}
+
+	return pqn->q->gws;
+}
+
+
 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
 {
 	struct kfd_process_device *pdd;
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

[PATCH 6/8] drm/amdgpu: Add function to add/remove gws to kfd process

[Zeng, Oak]

GWS bo is shared between all kfd processes. Add function to add gws
to kfd process's bo list so gws can be evicted from and restored
for process.

Change-Id: I75d74cfdadb7075ff8b2b68634e205deb73dc1ea
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       |   2 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 103 +++++++++++++++++++++--
 2 files changed, 100 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index c00c974..f968bf1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -155,6 +155,8 @@ int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
 void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj);
 int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size, void **mem_obj);
 void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj);
+int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem);
+int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem);
 uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
 				      enum kgd_engine_type type);
 void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index e1cae4a..02be141 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -457,6 +457,17 @@ static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
 	mutex_unlock(&process_info->lock);
 }
 
+static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
+		struct amdkfd_process_info *process_info)
+{
+	struct ttm_validate_buffer *bo_list_entry;
+
+	bo_list_entry = &mem->validate_list;
+	mutex_lock(&process_info->lock);
+	list_del(&bo_list_entry->head);
+	mutex_unlock(&process_info->lock);
+}
+
 /* Initializes user pages. It registers the MMU notifier and validates
  * the userptr BO in the GTT domain.
  *
@@ -1183,12 +1194,8 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 
 	if (user_addr) {
 		ret = init_user_pages(*mem, current->mm, user_addr);
-		if (ret) {
-			mutex_lock(&avm->process_info->lock);
-			list_del(&(*mem)->validate_list.head);
-			mutex_unlock(&avm->process_info->lock);
+		if (ret)
 			goto allocate_init_user_pages_failed;
-		}
 	}
 
 	if (offset)
@@ -1197,6 +1204,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
 	return 0;
 
 allocate_init_user_pages_failed:
+	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
 	amdgpu_bo_unref(&bo);
 	/* Don't unreserve system mem limit twice */
 	goto err_reserve_limit;
@@ -2104,3 +2112,88 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
 	kfree(pd_bo_list);
 	return ret;
 }
+
+int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
+{
+	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
+	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
+	int ret;
+
+	if (!info || !gws)
+		return -EINVAL;
+
+	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
+	if (!*mem)
+		return -EINVAL;
+
+	mutex_init(&(*mem)->lock);
+	(*mem)->bo = amdgpu_bo_ref(gws_bo);
+	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
+	(*mem)->process_info = process_info;
+	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
+	amdgpu_sync_create(&(*mem)->sync);
+
+
+	/* Validate gws bo the first time it is added to process */
+	mutex_lock(&(*mem)->process_info->lock);
+	ret = amdgpu_bo_reserve(gws_bo, false);
+	if (unlikely(ret)) {
+		pr_err("Reserve gws bo failed %d\n", ret);
+		goto bo_reservation_failure;
+	}
+
+	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, false);
+	if (ret) {
+		pr_err("GWS BO validate failed %d\n", ret);
+		goto bo_validation_failure;
+	}
+	/* GWS resource is shared b/t amdgpu and amdkfd
+	 * Add process eviction fence to bo so they can
+	 * evict each other.
+	 */
+	amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
+	amdgpu_bo_unreserve(gws_bo);
+	mutex_unlock(&(*mem)->process_info->lock);
+
+	return ret;
+
+bo_validation_failure:
+	amdgpu_bo_unreserve(gws_bo);
+bo_reservation_failure:
+	mutex_unlock(&(*mem)->process_info->lock);
+	amdgpu_sync_free(&(*mem)->sync);
+	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
+	amdgpu_bo_unref(&gws_bo);
+	mutex_destroy(&(*mem)->lock);
+	kfree(*mem);
+	*mem = NULL;
+	return ret;
+}
+
+int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
+{
+	int ret;
+	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
+	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
+	struct amdgpu_bo *gws_bo = kgd_mem->bo;
+
+	/* Remove BO from process's validate list so restore worker won't touch
+	 * it anymore
+	 */
+	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
+
+	ret = amdgpu_bo_reserve(gws_bo, false);
+	if (unlikely(ret)) {
+		pr_err("Reserve gws bo failed %d\n", ret);
+		//TODO add BO back to validate_list?
+		return ret;
+	}
+	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
+			process_info->eviction_fence);
+	amdgpu_bo_unreserve(gws_bo);
+	amdgpu_sync_free(&kgd_mem->sync);
+	amdgpu_bo_unref(&gws_bo);
+	mutex_destroy(&kgd_mem->lock);
+	kfree(mem);
+	return 0;
+}
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

[PATCH 7/8] drm/amdkfd: New IOCTL to allocate queue GWS

[Zeng, Oak]

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

Change-Id: I60153c26a577992ad873e4292e759e5c3d5bbd15
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c           | 39 ++++++++++++++++++++++
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  6 ++++
 include/uapi/linux/kfd_ioctl.h                     | 20 ++++++++++-
 3 files changed, 64 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 38ae53f..828bd66 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1559,6 +1559,43 @@ 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 kfd_dev *dev = NULL;
+	struct kgd_mem *mem;
+
+	if (args->num_gws == 0)
+		return -EINVAL;
+
+	if (!hws_gws_support ||
+		dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
+		return -EINVAL;
+
+	dev = kfd_device_by_id(args->gpu_id);
+	if (!dev) {
+		pr_debug("Could not find gpu id 0x%x\n", args->gpu_id);
+		return -EINVAL;
+	}
+
+	retval = amdgpu_amdkfd_add_gws_to_process(p->kgd_process_info,
+			dev->gws, &mem);
+	if (unlikely(retval))
+		return retval;
+
+	mutex_lock(&p->mutex);
+	retval = pqm_set_gws(&p->pqm, args->queue_id, mem);
+	mutex_unlock(&p->mutex);
+
+	if (unlikely(retval))
+		amdgpu_amdkfd_remove_gws_from_process(p->kgd_process_info, mem);
+
+	args->first_gws = 0;
+	return retval;
+}
+
 static int kfd_ioctl_get_dmabuf_info(struct file *filep,
 		struct kfd_process *p, void *data)
 {
@@ -1761,6 +1798,8 @@ 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)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
index 8e45296..e15ad48 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c
@@ -363,6 +363,12 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 		return -1;
 	}
 
+	if (pqm_get_gws(pqm, qid)) {
+		amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
+				pqm_get_gws(pqm, qid));
+		pqm_set_gws(pqm, qid, NULL);
+	}
+
 	if (pqn->kq) {
 		/* destroy kernel queue (DIQ) */
 		dqm = pqn->kq->dev->dqm;
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index 20917c5..912d690 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -410,6 +410,21 @@ struct kfd_ioctl_unmap_memory_from_gpu_args {
 	__u32 n_success;		/* to/from KFD */
 };
 
+/* Allocate GWS for specific queue
+ *
+ * @gpu_id:      device identifier
+ * @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 gpu_id;		/* to KFD */
+	__u32 queue_id;		/* to KFD */
+	__u32 num_gws;		/* to KFD */
+	__u32 first_gws;	/* from KFD */
+};
+
 struct kfd_ioctl_get_dmabuf_info_args {
 	__u64 size;		/* from KFD */
 	__u64 metadata_ptr;	/* to KFD */
@@ -529,7 +544,10 @@ 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(0x22, struct kfd_ioctl_alloc_queue_gws_args)
+
 #define AMDKFD_COMMAND_START		0x01
-#define AMDKFD_COMMAND_END		0x1E
+#define AMDKFD_COMMAND_END		0x23
 
 #endif
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx