[PATCH 0/2] add API to get GPU physical address

[PATCH 0/2] add API to get GPU physical address

[Ajit Khaparde]

Add API to get the physical address of the peer GPU.
This should allow some NIC hardware to directly use the
physical address for DMA instead of the CUDA Unified Memory
provided by current mem_map calls.

A dependent code change has been pushed to the gdrcopy tree
on github. The pull request is open here.
https://github.com/NVIDIA/gdrcopy/pull/304

Please review.

Thanks

Ajit Khaparde (2):
  gpudev: add API to get GPU physical address
  gpu/cuda: extend cuda code to get PA of GPU

 drivers/gpu/cuda/common.h  |   2 +
 drivers/gpu/cuda/cuda.c    | 117 ++++++++++++++++++++++++++++++-------
 drivers/gpu/cuda/gdrcopy.c |  50 ++++++++++++++++
 lib/gpudev/gpudev.c        |  61 +++++++++++++++++++
 lib/gpudev/gpudev_driver.h |   6 ++
 lib/gpudev/rte_gpudev.h    |  49 ++++++++++++++++
 lib/gpudev/version.map     |   2 +
 7 files changed, 265 insertions(+), 22 deletions(-)

-- 
2.39.3 (Apple Git-146)

[PATCH 1/2] gpudev: add API to get GPU physical address

[Ajit Khaparde]

Add API to get the physical address of the peer GPU.
This should allow some NIC hardware to directly use the
physical address for DMA instead of the CUDA Unified Memory
provided by rte_gpu_mem_map.

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 lib/gpudev/gpudev.c        | 61 ++++++++++++++++++++++++++++++++++++++
 lib/gpudev/gpudev_driver.h |  6 ++++
 lib/gpudev/rte_gpudev.h    | 49 ++++++++++++++++++++++++++++++
 lib/gpudev/version.map     |  2 ++
 4 files changed, 118 insertions(+)

diff --git a/lib/gpudev/gpudev.c b/lib/gpudev/gpudev.c
index 1c2011b856..acbe39361b 100644
--- a/lib/gpudev/gpudev.c
+++ b/lib/gpudev/gpudev.c
@@ -683,6 +683,44 @@ rte_gpu_mem_cpu_map(int16_t dev_id, size_t size, void *ptr)
 	}
 }
 
+void *
+rte_gpu_mem_dma_map(int16_t dev_id, size_t size, void *ptr)
+{
+	struct rte_gpu *dev;
+	void *ptr_out;
+	int ret;
+
+	dev = gpu_get_by_id(dev_id);
+	if (dev == NULL) {
+		GPU_LOG(ERR, "mem CPU map for invalid device ID %d", dev_id);
+		rte_errno = ENODEV;
+		return NULL;
+	}
+
+	if (dev->ops.mem_cpu_map == NULL) {
+		GPU_LOG(ERR, "mem CPU map not supported");
+		rte_errno = ENOTSUP;
+		return NULL;
+	}
+
+	if (ptr == NULL || size == 0) /* dry-run  */
+		return NULL;
+
+	ret = GPU_DRV_RET(dev->ops.mem_dma_map(dev, size, ptr, &ptr_out));
+
+	switch (ret) {
+	case 0:
+		return ptr_out;
+	case -ENOMEM:
+	case -E2BIG:
+		rte_errno = -ret;
+		return NULL;
+	default:
+		rte_errno = -EPERM;
+		return NULL;
+	}
+}
+
 int
 rte_gpu_mem_cpu_unmap(int16_t dev_id, void *ptr)
 {
@@ -706,6 +744,29 @@ rte_gpu_mem_cpu_unmap(int16_t dev_id, void *ptr)
 	return GPU_DRV_RET(dev->ops.mem_cpu_unmap(dev, ptr));
 }
 
+int
+rte_gpu_mem_dma_unmap(int16_t dev_id, void *ptr)
+{
+	struct rte_gpu *dev;
+
+	dev = gpu_get_by_id(dev_id);
+	if (dev == NULL) {
+		GPU_LOG(ERR, "cpu_unmap mem for invalid device ID %d", dev_id);
+		rte_errno = ENODEV;
+		return -rte_errno;
+	}
+
+	if (dev->ops.mem_dma_unmap == NULL) {
+		rte_errno = ENOTSUP;
+		return -rte_errno;
+	}
+
+	if (ptr == NULL) /* dry-run */
+		return 0;
+
+	return GPU_DRV_RET(dev->ops.mem_dma_unmap(dev, ptr));
+}
+
 int
 rte_gpu_wmb(int16_t dev_id)
 {
diff --git a/lib/gpudev/gpudev_driver.h b/lib/gpudev/gpudev_driver.h
index 37b6ae3149..ad3ab9e214 100644
--- a/lib/gpudev/gpudev_driver.h
+++ b/lib/gpudev/gpudev_driver.h
@@ -39,6 +39,8 @@ typedef int (rte_gpu_mem_unregister_t)(struct rte_gpu *dev, void *ptr);
 typedef int (rte_gpu_mem_cpu_map_t)(struct rte_gpu *dev, size_t size, void *ptr_in, void **ptr_out);
 typedef int (rte_gpu_mem_cpu_unmap_t)(struct rte_gpu *dev, void *ptr);
 typedef int (rte_gpu_wmb_t)(struct rte_gpu *dev);
+typedef int (rte_gpu_mem_dma_map_t)(struct rte_gpu *dev, size_t size, void *ptr_in, void **ptr_out);
+typedef int (rte_gpu_mem_dma_unmap_t)(struct rte_gpu *dev, void *ptr);
 
 struct rte_gpu_ops {
 	/* Get device info. If NULL, info is just copied. */
@@ -59,6 +61,10 @@ struct rte_gpu_ops {
 	rte_gpu_mem_cpu_unmap_t *mem_cpu_unmap;
 	/* Enforce GPU write memory barrier. */
 	rte_gpu_wmb_t *wmb;
+	/* DMA address of GPU memory. */
+	rte_gpu_mem_dma_map_t *mem_dma_map;
+	/* DMA unmap GPU memory. */
+	rte_gpu_mem_dma_unmap_t *mem_dma_unmap;
 };
 
 struct rte_gpu_mpshared {
diff --git a/lib/gpudev/rte_gpudev.h b/lib/gpudev/rte_gpudev.h
index 0a94a6abc4..ebc789880b 100644
--- a/lib/gpudev/rte_gpudev.h
+++ b/lib/gpudev/rte_gpudev.h
@@ -484,6 +484,34 @@ int rte_gpu_mem_unregister(int16_t dev_id, void *ptr);
 __rte_experimental
 void *rte_gpu_mem_cpu_map(int16_t dev_id, size_t size, void *ptr);
 
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * Map a chunk of GPU memory to make it accessible for DMA
+ * using the memory pointer returned by the function.
+ * GPU memory has to be allocated via rte_gpu_mem_alloc().
+ *
+ * @param dev_id
+ *   Device ID requiring mapped memory.
+ * @param size
+ *   Number of bytes to map.
+ *   Requesting 0 will do nothing.
+ * @param ptr
+ *   Pointer to the GPU memory area to be mapped for DMA.
+ *   NULL is a no-op accepted value.
+
+ * @return
+ *   A pointer to the mapped GPU memory usable by the CPU, otherwise NULL and rte_errno is set:
+ *   - ENODEV if invalid dev_id
+ *   - ENOTSUP if operation not supported by the driver
+ *   - E2BIG if size is higher than limit
+ *   - ENOMEM if out of space
+ *   - EPERM if driver error
+ */
+__rte_experimental
+void *rte_gpu_mem_dma_map(int16_t dev_id, size_t size, void *ptr);
+
 /**
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice.
@@ -505,6 +533,27 @@ void *rte_gpu_mem_cpu_map(int16_t dev_id, size_t size, void *ptr);
 __rte_experimental
 int rte_gpu_mem_cpu_unmap(int16_t dev_id, void *ptr);
 
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * Unmap a chunk of GPU memory previously mapped with rte_gpu_mem_dma_map()
+ *
+ * @param dev_id
+ *   Reference device ID.
+ * @param ptr
+ *   Pointer to the GPU memory area to be unmapped.
+ *   NULL is a no-op accepted value.
+ *
+ * @return
+ *   0 on success, -rte_errno otherwise:
+ *   - ENODEV if invalid dev_id
+ *   - ENOTSUP if operation not supported by the driver
+ *   - EPERM if driver error
+ */
+__rte_experimental
+int rte_gpu_mem_dma_unmap(int16_t dev_id, void *ptr);
+
 /**
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice.
diff --git a/lib/gpudev/version.map b/lib/gpudev/version.map
index a2c8ce5759..f267b5e029 100644
--- a/lib/gpudev/version.map
+++ b/lib/gpudev/version.map
@@ -28,6 +28,8 @@ EXPERIMENTAL {
 	rte_gpu_mem_cpu_unmap;
 	rte_gpu_mem_unregister;
 	rte_gpu_wmb;
+	rte_gpu_mem_dma_map;
+	rte_gpu_mem_dma_unmap;
 };
 
 INTERNAL {
-- 
2.39.3 (Apple Git-146)

[PATCH 2/2] gpu/cuda: extend cuda code to get PA of GPU

[Ajit Khaparde]

Add code to get the physical address of the GPU memory.
This should allow some NIC hardware to directly use the
physical address for DMA instead of the CUDA Unified Memory
provided by cuda_mem_map.

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 drivers/gpu/cuda/common.h  |   2 +
 drivers/gpu/cuda/cuda.c    | 117 ++++++++++++++++++++++++++++++-------
 drivers/gpu/cuda/gdrcopy.c |  50 ++++++++++++++++
 3 files changed, 147 insertions(+), 22 deletions(-)

diff --git a/drivers/gpu/cuda/common.h b/drivers/gpu/cuda/common.h
index e809988503..5da5e7d176 100644
--- a/drivers/gpu/cuda/common.h
+++ b/drivers/gpu/cuda/common.h
@@ -33,6 +33,8 @@ extern int cuda_logtype;
 
 int gdrcopy_pin(gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
 		uint64_t d_addr, size_t size, void **h_addr);
+int gdrcopy_dma(gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
+		uint64_t d_addr, size_t size, void **h_addr, uint64_t *paddr);
 int gdrcopy_unpin(gdr_t gdrc_h, __rte_unused gdr_mh_t mh,
 		void *d_addr, size_t size);
 
diff --git a/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
index a552aabeb8..36cff7b3a5 100644
--- a/drivers/gpu/cuda/cuda.c
+++ b/drivers/gpu/cuda/cuda.c
@@ -959,6 +959,87 @@ cuda_mem_cpu_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in, vo
 	return 0;
 }
 
+static int
+cuda_mem_dma_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in, void **ptr_out)
+{
+	struct mem_entry *mem_item;
+	cuda_ptr_key hk;
+	uint64_t paddr;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	hk = get_hash_from_ptr((void *)ptr_in);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory.", ptr_in);
+		rte_errno = EPERM;
+		return -rte_errno;
+	}
+
+	if (mem_item->mtype != GPU_MEM) {
+		rte_cuda_log(ERR, "Memory address 0x%p is not GPU memory type.", ptr_in);
+		rte_errno = EPERM;
+		return -rte_errno;
+	}
+
+	if (mem_item->size != size)
+		rte_cuda_log(WARNING,
+				"Can't expose memory area with size (%zd) different from original size (%zd).",
+				size, mem_item->size);
+
+	if (gdrcopy_dma(&gdrc_h, &(mem_item->mh), (uint64_t)mem_item->ptr_d,
+					mem_item->size, &(mem_item->ptr_h), &paddr)) {
+		rte_cuda_log(ERR, "Error exposing GPU memory address 0x%p.", ptr_in);
+		rte_errno = EPERM;
+		return -rte_errno;
+	}
+
+	mem_item->mtype = GPU_REGISTERED;
+	*ptr_out = (void *)paddr;
+
+	return 0;
+}
+
+static int
+cuda_mem_free(struct rte_gpu *dev, void *ptr)
+{
+	CUresult res;
+	struct mem_entry *mem_item;
+	const char *err_string;
+	cuda_ptr_key hk;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	hk = get_hash_from_ptr((void *)ptr);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
+		rte_errno = EPERM;
+		return -rte_errno;
+	}
+
+	if (mem_item->mtype == GPU_MEM) {
+		res = pfn_cuMemFree(mem_item->ptr_orig_d);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuMemFree current failed with %s",
+					err_string);
+			rte_errno = EPERM;
+			return -rte_errno;
+		}
+
+		return mem_list_del_item(hk);
+	}
+
+	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
+
+	return -EPERM;
+}
+
 static int
 cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
 {
@@ -1034,48 +1115,38 @@ cuda_mem_cpu_unmap(struct rte_gpu *dev, void *ptr_in)
 }
 
 static int
-cuda_mem_free(struct rte_gpu *dev, void *ptr)
+cuda_mem_dma_unmap(struct rte_gpu *dev, void *ptr_in)
 {
-	CUresult res;
 	struct mem_entry *mem_item;
-	const char *err_string;
 	cuda_ptr_key hk;
 
 	if (dev == NULL)
 		return -ENODEV;
 
-	hk = get_hash_from_ptr((void *)ptr);
+	hk = get_hash_from_ptr((void *)ptr_in);
 
 	mem_item = mem_list_find_item(hk);
 	if (mem_item == NULL) {
-		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
+		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory.", ptr_in);
 		rte_errno = EPERM;
 		return -rte_errno;
 	}
 
-	/*
-	 * If a GPU memory area that's CPU mapped is being freed
-	 * without calling cpu_unmap, force the unmapping.
-	 */
-	if (mem_item->mtype == GPU_REGISTERED)
-		cuda_mem_cpu_unmap(dev, ptr);
-
-	if (mem_item->mtype == GPU_MEM) {
-		res = pfn_cuMemFree(mem_item->ptr_orig_d);
-		if (res != 0) {
-			pfn_cuGetErrorString(res, &(err_string));
-			rte_cuda_log(ERR, "cuMemFree current failed with %s",
-					err_string);
+	if (mem_item->mtype == GPU_REGISTERED) {
+		if (gdrcopy_unpin(gdrc_h, mem_item->mh, (void *)mem_item->ptr_d,
+				mem_item->size)) {
+			rte_cuda_log(ERR, "Error unexposing GPU memory address 0x%p.", ptr_in);
 			rte_errno = EPERM;
 			return -rte_errno;
 		}
 
-		return mem_list_del_item(hk);
+		mem_item->mtype = GPU_MEM;
+	} else {
+		rte_errno = EPERM;
+		return -rte_errno;
 	}
 
-	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
-
-	return -EPERM;
+	return 0;
 }
 
 static int
@@ -1391,7 +1462,9 @@ cuda_gpu_probe(__rte_unused struct rte_pci_driver *pci_drv, struct rte_pci_devic
 	dev->ops.mem_register = cuda_mem_register;
 	dev->ops.mem_unregister = cuda_mem_unregister;
 	dev->ops.mem_cpu_map = cuda_mem_cpu_map;
+	dev->ops.mem_dma_map = cuda_mem_dma_map;
 	dev->ops.mem_cpu_unmap = cuda_mem_cpu_unmap;
+	dev->ops.mem_dma_unmap = cuda_mem_dma_unmap;
 	dev->ops.wmb = cuda_wmb;
 
 	rte_gpu_complete_new(dev);
diff --git a/drivers/gpu/cuda/gdrcopy.c b/drivers/gpu/cuda/gdrcopy.c
index bd56b73ce4..ac6b3d773a 100644
--- a/drivers/gpu/cuda/gdrcopy.c
+++ b/drivers/gpu/cuda/gdrcopy.c
@@ -12,6 +12,8 @@ static void *gdrclib;
 static gdr_t (*sym_gdr_open)(void);
 static int (*sym_gdr_pin_buffer)(gdr_t g, unsigned long addr, size_t size,
 		uint64_t p2p_token, uint32_t va_space, gdr_mh_t *handle);
+static int (*sym_gdr_p2p_dma_map_buffer)(gdr_t g, unsigned long addr, size_t size,
+		uint64_t p2p_token, uint32_t va_space, gdr_mh_t *handle, uint64_t *paddr);
 static int (*sym_gdr_unpin_buffer)(gdr_t g, gdr_mh_t handle);
 static int (*sym_gdr_map)(gdr_t g, gdr_mh_t handle, void **va, size_t size);
 static int (*sym_gdr_unmap)(gdr_t g, gdr_mh_t handle, void *va, size_t size);
@@ -45,6 +47,13 @@ gdrcopy_loader(void)
 		return -1;
 	}
 
+	sym_gdr_p2p_dma_map_buffer = dlsym(gdrclib, "gdr_p2p_dma_map_buffer");
+	if (sym_gdr_p2p_dma_map_buffer == NULL) {
+		rte_cuda_log(ERR, "Failed to load GDRCopy symbol gdr_p2p_dma_map_buffer\n");
+		printf("Failed to load GDRCopy symbol gdr_p2p_dma_map_buffer\n");
+		return -1;
+	}
+
 	sym_gdr_unpin_buffer = dlsym(gdrclib, "gdr_unpin_buffer");
 	if (sym_gdr_unpin_buffer == NULL) {
 		rte_cuda_log(ERR, "Failed to load GDRCopy symbols\n");
@@ -119,6 +128,47 @@ gdrcopy_pin(__rte_unused gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
 #endif
 }
 
+int
+gdrcopy_dma(__rte_unused gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
+		__rte_unused uint64_t d_addr, __rte_unused size_t size,
+		__rte_unused void **h_addr, __rte_unused uint64_t *paddr)
+{
+#ifdef DRIVERS_GPU_CUDA_GDRCOPY_H
+	uint64_t phys;
+
+	if (*gdrc_h == NULL) {
+		if (gdrcopy_loader())
+			return -ENOTSUP;
+
+		if (gdrcopy_open(gdrc_h)) {
+			rte_cuda_log(ERR,
+					"GDRCopy gdrdrv kernel module not found. Can't CPU map GPU memory.");
+			return -EPERM;
+		}
+	}
+
+	/* Pin the device buffer */
+	if (sym_gdr_p2p_dma_map_buffer(*gdrc_h, d_addr, size, 0, 0, mh, &phys) != 0) {
+		rte_cuda_log(ERR, "GDRCopy p2p dma map buffer error.");
+		return -1;
+	}
+	*paddr = phys;
+
+	/* Map the buffer to user space */
+	if (sym_gdr_map(*gdrc_h, *mh, h_addr, size) != 0) {
+		rte_cuda_log(ERR, "GDRCopy map buffer error.");
+		sym_gdr_unpin_buffer(*gdrc_h, *mh);
+		return -1;
+	}
+
+	return 0;
+#else
+	rte_cuda_log(ERR,
+			"GDRCopy headers not provided at DPDK building time. Can't CPU map GPU memory.");
+	return -ENOTSUP;
+#endif
+}
+
 int
 gdrcopy_unpin(gdr_t gdrc_h, __rte_unused gdr_mh_t mh,
 		__rte_unused void *d_addr, __rte_unused size_t size)
-- 
2.39.3 (Apple Git-146)

Re: [PATCH 2/2] gpu/cuda: extend cuda code to get PA of GPU

[Stephen Hemminger]

On Thu,  5 Sep 2024 21:16:14 -0700
Ajit Khaparde <ajit.khaparde@broadcom.com> wrote:

> +	sym_gdr_p2p_dma_map_buffer = dlsym(gdrclib, "gdr_p2p_dma_map_buffer");
> +	if (sym_gdr_p2p_dma_map_buffer == NULL) {
> +		rte_cuda_log(ERR, "Failed to load GDRCopy symbol gdr_p2p_dma_map_buffer\n");
> +		printf("Failed to load GDRCopy symbol gdr_p2p_dma_map_buffer\n");
> +		return -1;
> +	}

No calls to printf() are acceptable in DPDK drivers.
There is a log message, that is sufficient here.