diff mbox series

[02/15] net: hbl_cn: memory manager component

Message ID 20240613082208.1439968-3-oshpigelman@habana.ai (mailing list archive)
State New, archived
Headers show
Series Introduce HabanaLabs network drivers | expand

Commit Message

Omer Shpigelman June 13, 2024, 8:21 a.m. UTC
Add a common memory manager which handles allocation and mapping.
It manages physical/virtual memory on host/device.

Signed-off-by: Omer Shpigelman <oshpigelman@habana.ai>
Co-developed-by: Abhilash K V <kvabhilash@habana.ai>
Signed-off-by: Abhilash K V <kvabhilash@habana.ai>
Co-developed-by: Andrey Agranovich <aagranovich@habana.ai>
Signed-off-by: Andrey Agranovich <aagranovich@habana.ai>
Co-developed-by: Bharat Jauhari <bjauhari@habana.ai>
Signed-off-by: Bharat Jauhari <bjauhari@habana.ai>
Co-developed-by: David Meriin <dmeriin@habana.ai>
Signed-off-by: David Meriin <dmeriin@habana.ai>
Co-developed-by: Sagiv Ozeri <sozeri@habana.ai>
Signed-off-by: Sagiv Ozeri <sozeri@habana.ai>
Co-developed-by: Zvika Yehudai <zyehudai@habana.ai>
Signed-off-by: Zvika Yehudai <zyehudai@habana.ai>
---
 .../intel/hbl_cn/common/hbl_cn_memory.c       | 325 +++++++++++++++++-
 1 file changed, 322 insertions(+), 3 deletions(-)
diff mbox series

Patch

diff --git a/drivers/net/ethernet/intel/hbl_cn/common/hbl_cn_memory.c b/drivers/net/ethernet/intel/hbl_cn/common/hbl_cn_memory.c
index 93c97fad6a20..878ecba66aa3 100644
--- a/drivers/net/ethernet/intel/hbl_cn/common/hbl_cn_memory.c
+++ b/drivers/net/ethernet/intel/hbl_cn/common/hbl_cn_memory.c
@@ -4,37 +4,356 @@ 
  * All Rights Reserved.
  */
 
+#include <linux/vmalloc.h>
 #include "hbl_cn.h"
 
-int hbl_cn_mem_alloc(struct hbl_cn_ctx *ctx, struct hbl_cn_mem_data *mem_data)
+static int hbl_cn_map_vmalloc_range(struct hbl_cn_ctx *ctx, u64 vmalloc_va, u64 device_va,
+				    u64 size)
+{
+	struct hbl_cn_aux_ops *aux_ops;
+	struct hbl_aux_dev *aux_dev;
+
+	aux_dev = ctx->hdev->cn_aux_dev;
+	aux_ops = aux_dev->aux_ops;
+
+	return aux_ops->vm_dev_mmu_map(aux_dev, ctx->driver_vm_info.vm_handle, HBL_CN_MEM_TYPE_HOST,
+				       vmalloc_va, device_va, size);
+}
+
+static void hbl_cn_unmap_vmalloc_range(struct hbl_cn_ctx *ctx, u64 device_va, u64 size)
+{
+	struct hbl_cn_aux_ops *aux_ops;
+	struct hbl_aux_dev *aux_dev;
+
+	aux_dev = ctx->hdev->cn_aux_dev;
+	aux_ops = aux_dev->aux_ops;
+
+	aux_ops->vm_dev_mmu_unmap(aux_dev, ctx->driver_vm_info.vm_handle, device_va, size);
+}
+
+static int alloc_mem(struct hbl_cn_mem_buf *buf, gfp_t gfp, struct hbl_cn_mem_data *mem_data)
+{
+	u64 device_addr, size = mem_data->size;
+	struct hbl_cn_ctx *ctx = buf->ctx;
+	u32 mem_id = mem_data->mem_id;
+	struct hbl_cn_device *hdev;
+	void *p = NULL;
+
+	hdev = ctx->hdev;
+
+	switch (mem_id) {
+	case HBL_CN_DRV_MEM_HOST_DMA_COHERENT:
+		if (get_order(size) > MAX_PAGE_ORDER) {
+			dev_err(hdev->dev, "memory size 0x%llx must be less than 0x%lx\n", size,
+				1UL << (PAGE_SHIFT + MAX_PAGE_ORDER));
+			return -ENOMEM;
+		}
+
+		p = hbl_cn_dma_alloc_coherent(hdev, size, &buf->bus_address, GFP_USER | __GFP_ZERO);
+		if (!p) {
+			dev_err(hdev->dev,
+				"failed to allocate 0x%llx of dma memory for the NIC\n", size);
+			return -ENOMEM;
+		}
+
+		break;
+	case HBL_CN_DRV_MEM_HOST_VIRTUAL:
+		p = vmalloc_user(size);
+		if (!p) {
+			dev_err(hdev->dev, "failed to allocate vmalloc memory, size 0x%llx\n",
+				size);
+			return -ENOMEM;
+		}
+
+		break;
+	case HBL_CN_DRV_MEM_HOST_MAP_ONLY:
+		p = mem_data->in.host_map_data.kernel_address;
+		buf->bus_address = mem_data->in.host_map_data.bus_address;
+		break;
+	case HBL_CN_DRV_MEM_DEVICE:
+		if (!hdev->wq_arrays_pool_enable) {
+			dev_err(hdev->dev, "No WQ arrays pool support for device memory\n");
+			return -EOPNOTSUPP;
+		}
+
+		device_addr = (u64)gen_pool_alloc(hdev->wq_arrays_pool, size);
+		if (!device_addr) {
+			dev_err(hdev->dev, "Failed to allocate device memory, size 0x%llx\n", size);
+			return -ENOMEM;
+		}
+
+		buf->device_addr = device_addr;
+		break;
+	default:
+		dev_err(hdev->dev, "Invalid mem_id %d\n", mem_id);
+		return -EINVAL;
+	}
+
+	buf->kernel_address = p;
+	buf->mappable_size = size;
+
+	return 0;
+}
+
+static int map_mem(struct hbl_cn_mem_buf *buf, struct hbl_cn_mem_data *mem_data)
+{
+	struct hbl_cn_ctx *ctx = buf->ctx;
+	struct hbl_cn_device *hdev;
+	int rc;
+
+	hdev = ctx->hdev;
+
+	if (mem_data->mem_id == HBL_CN_DRV_MEM_HOST_DMA_COHERENT) {
+		dev_err(hdev->dev, "Mapping DMA coherent host memory is not yet supported\n");
+		return -EPERM;
+	}
+
+	rc = hbl_cn_map_vmalloc_range(ctx, (u64)buf->kernel_address, mem_data->device_va,
+				      buf->mappable_size);
+	if (rc)
+		return rc;
+
+	buf->device_va = mem_data->device_va;
+
+	return 0;
+}
+
+static void mem_do_release(struct hbl_cn_device *hdev, struct hbl_cn_mem_buf *buf)
+{
+	struct hbl_cn_asic_funcs *asic_funcs = hdev->asic_funcs;
+
+	if (buf->mem_id == HBL_CN_DRV_MEM_HOST_DMA_COHERENT)
+		asic_funcs->dma_free_coherent(hdev, buf->mappable_size, buf->kernel_address,
+					      buf->bus_address);
+	else if (buf->mem_id == HBL_CN_DRV_MEM_HOST_VIRTUAL)
+		vfree(buf->kernel_address);
+	else if (buf->mem_id == HBL_CN_DRV_MEM_DEVICE)
+		gen_pool_free(hdev->wq_arrays_pool, buf->device_addr, buf->mappable_size);
+}
+
+static int __cn_mem_buf_alloc(struct hbl_cn_mem_buf *buf, gfp_t gfp,
+			      struct hbl_cn_mem_data *mem_data)
+{
+	struct hbl_cn_ctx *ctx = buf->ctx;
+	struct hbl_cn_device *hdev;
+	int rc;
+
+	hdev = ctx->hdev;
+
+	if (mem_data->mem_id != HBL_CN_DRV_MEM_DEVICE)
+		mem_data->size = PAGE_ALIGN(mem_data->size);
+
+	rc = alloc_mem(buf, gfp, mem_data);
+	if (rc)
+		return rc;
+
+	if (mem_data->device_va) {
+		mem_data->device_va = PAGE_ALIGN(mem_data->device_va);
+		rc = map_mem(buf, mem_data);
+		if (rc)
+			goto release_mem;
+	}
+
+	return 0;
+
+release_mem:
+	mem_do_release(hdev, buf);
+	return rc;
+}
+
+static struct hbl_cn_mem_buf *cn_mem_buf_alloc(struct hbl_cn_ctx *ctx, gfp_t gfp,
+					       struct hbl_cn_mem_data *mem_data)
+{
+	struct xa_limit id_limit = XA_LIMIT(1, INT_MAX);
+	struct hbl_cn_device *hdev = ctx->hdev;
+	struct hbl_cn_mem_buf *buf;
+	int rc;
+	u32 id;
+
+	buf = kzalloc(sizeof(*buf), gfp);
+	if (!buf)
+		return NULL;
+
+	rc = xa_alloc(&hdev->mem_ids, &id, buf, id_limit, GFP_ATOMIC);
+	if (rc) {
+		dev_err(hdev->dev, "Failed to allocate xarray for a new buffer, rc=%d\n", rc);
+		goto free_buf;
+	}
+
+	buf->ctx = ctx;
+	buf->mem_id = mem_data->mem_id;
+
+	buf->handle = (((u64)id | hdev->mmap_type_flag) << PAGE_SHIFT);
+	kref_init(&buf->refcount);
+
+	rc = __cn_mem_buf_alloc(buf, gfp, mem_data);
+	if (rc)
+		goto remove_xa;
+
+	return buf;
+
+remove_xa:
+	xa_erase(&hdev->mem_ids, lower_32_bits(buf->handle >> PAGE_SHIFT));
+free_buf:
+	kfree(buf);
+	return NULL;
+}
+
+static int cn_mem_alloc(struct hbl_cn_ctx *ctx, struct hbl_cn_mem_data *mem_data)
 {
+	struct hbl_cn_mem_buf *buf;
+
+	buf = cn_mem_buf_alloc(ctx, GFP_KERNEL, mem_data);
+	if (!buf)
+		return -ENOMEM;
+
+	mem_data->handle = buf->handle;
+
+	if (mem_data->mem_id == HBL_CN_DRV_MEM_HOST_DMA_COHERENT)
+		mem_data->addr = (u64)buf->bus_address;
+	else if (mem_data->mem_id == HBL_CN_DRV_MEM_HOST_VIRTUAL)
+		mem_data->addr = (u64)buf->kernel_address;
+	else if (mem_data->mem_id == HBL_CN_DRV_MEM_DEVICE)
+		mem_data->addr = (u64)buf->device_addr;
+
 	return 0;
 }
 
+int hbl_cn_mem_alloc(struct hbl_cn_ctx *ctx, struct hbl_cn_mem_data *mem_data)
+{
+	struct hbl_cn_device *hdev = ctx->hdev;
+	int rc;
+
+	switch (mem_data->mem_id) {
+	case HBL_CN_DRV_MEM_HOST_DMA_COHERENT:
+	case HBL_CN_DRV_MEM_HOST_VIRTUAL:
+	case HBL_CN_DRV_MEM_HOST_MAP_ONLY:
+	case HBL_CN_DRV_MEM_DEVICE:
+		rc = cn_mem_alloc(ctx, mem_data);
+		break;
+	default:
+		dev_dbg(hdev->dev, "Invalid mem_id %d\n", mem_data->mem_id);
+		rc = -EINVAL;
+		break;
+	}
+
+	return rc;
+}
+
+static void cn_mem_buf_destroy(struct hbl_cn_mem_buf *buf)
+{
+	if (buf->device_va)
+		hbl_cn_unmap_vmalloc_range(buf->ctx, buf->device_va, buf->mappable_size);
+
+	mem_do_release(buf->ctx->hdev, buf);
+
+	kfree(buf);
+}
+
 int hbl_cn_mem_destroy(struct hbl_cn_device *hdev, u64 handle)
 {
+	struct hbl_cn_mem_buf *buf;
+	int rc;
+
+	buf = hbl_cn_mem_buf_get(hdev, handle);
+	if (!buf) {
+		dev_dbg(hdev->dev, "Memory destroy failed, no match for handle 0x%llx\n", handle);
+		return -EINVAL;
+	}
+
+	rc = atomic_cmpxchg(&buf->is_destroyed, 0, 1);
+	hbl_cn_mem_buf_put(buf);
+	if (rc) {
+		dev_dbg(hdev->dev, "Memory destroy failed, handle 0x%llx was already destroyed\n",
+			handle);
+		return -EINVAL;
+	}
+
+	rc = hbl_cn_mem_buf_put_handle(hdev, handle);
+	if (rc < 0)
+		return rc;
+
+	if (rc == 0)
+		dev_dbg(hdev->dev, "Handle 0x%llx is destroyed while still in use\n", handle);
+
 	return 0;
 }
 
+static void cn_mem_buf_release(struct kref *kref)
+{
+	struct hbl_cn_mem_buf *buf = container_of(kref, struct hbl_cn_mem_buf, refcount);
+	struct hbl_cn_device *hdev = buf->ctx->hdev;
+
+	xa_erase(&hdev->mem_ids, lower_32_bits(buf->handle >> PAGE_SHIFT));
+
+	cn_mem_buf_destroy(buf);
+}
+
 struct hbl_cn_mem_buf *hbl_cn_mem_buf_get(struct hbl_cn_device *hdev, u64 handle)
 {
-	return NULL;
+	struct hbl_cn_mem_buf *buf;
+
+	xa_lock(&hdev->mem_ids);
+	buf = xa_load(&hdev->mem_ids, lower_32_bits(handle >> PAGE_SHIFT));
+	if (!buf) {
+		xa_unlock(&hdev->mem_ids);
+		dev_dbg(hdev->dev, "Buff get failed, no match to handle %#llx\n", handle);
+		return NULL;
+	}
+
+	kref_get(&buf->refcount);
+	xa_unlock(&hdev->mem_ids);
+
+	return buf;
 }
 
 int hbl_cn_mem_buf_put(struct hbl_cn_mem_buf *buf)
 {
-	return 0;
+	return kref_put(&buf->refcount, cn_mem_buf_release);
+}
+
+static void cn_mem_buf_remove_xa_locked(struct kref *kref)
+{
+	struct hbl_cn_mem_buf *buf = container_of(kref, struct hbl_cn_mem_buf, refcount);
+
+	__xa_erase(&buf->ctx->hdev->mem_ids, lower_32_bits(buf->handle >> PAGE_SHIFT));
 }
 
 int hbl_cn_mem_buf_put_handle(struct hbl_cn_device *hdev, u64 handle)
 {
+	struct hbl_cn_mem_buf *buf;
+
+	xa_lock(&hdev->mem_ids);
+	buf = xa_load(&hdev->mem_ids, lower_32_bits(handle >> PAGE_SHIFT));
+	if (!buf) {
+		xa_unlock(&hdev->mem_ids);
+		dev_dbg(hdev->dev, "Buff put failed, no match to handle %#llx\n", handle);
+		return -EINVAL;
+	}
+
+	if (kref_put(&buf->refcount, cn_mem_buf_remove_xa_locked)) {
+		xa_unlock(&hdev->mem_ids);
+		cn_mem_buf_destroy(buf);
+		return 1;
+	}
+
+	xa_unlock(&hdev->mem_ids);
 	return 0;
 }
 
 void hbl_cn_mem_init(struct hbl_cn_device *hdev)
 {
+	xa_init_flags(&hdev->mem_ids, XA_FLAGS_ALLOC);
 }
 
 void hbl_cn_mem_fini(struct hbl_cn_device *hdev)
 {
+	struct xarray *mem_ids;
+
+	mem_ids = &hdev->mem_ids;
+
+	if (!xa_empty(mem_ids))
+		dev_crit(hdev->dev, "memory manager is destroyed while not empty!\n");
+
+	xa_destroy(mem_ids);
 }