1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright (c) 2021 NVIDIA Corporation & Affiliates
4 General-Purpose Graphics Processing Unit Library
5 ================================================
7 When mixing networking activity with task processing on a GPU device,
8 there may be the need to put in communication the CPU with the device
9 in order to manage the memory, synchronize operations, exchange info, etc..
11 By means of the generic GPU interface provided by this library,
12 it is possible to allocate a chunk of GPU memory and use it
13 to create a DPDK mempool with external mbufs having the payload
14 on the GPU memory, enabling any network interface card
15 (which support this feature like Mellanox NIC)
16 to directly transmit and receive packets using GPU memory.
18 Additionally, this library provides a number of functions
19 to enhance the dialog between CPU and GPU.
21 Out of scope of this library is to provide a wrapper for GPU specific libraries
22 (e.g. CUDA Toolkit or OpenCL), thus it is not possible to launch workload
23 on the device or create GPU specific objects
24 (e.g. CUDA Driver context or CUDA Streams in case of NVIDIA GPUs).
26 This library is optional in DPDK
27 and can be disabled with ``-Ddisable_libs=gpudev``.
33 This library provides a number of features:
35 - Interoperability with device-specific library through generic handlers.
36 - Allocate and free memory on the device.
37 - Register CPU memory to make it visible from the device.
38 - Communication between the CPU and the device.
40 The whole CPU - GPU communication is implemented
41 using CPU memory visible from the GPU.
50 By default, DPDK PCIe module detects and registers physical GPU devices
52 With the gpudev library is also possible to add additional non-physical devices
53 through an ``uint64_t`` generic handler (e.g. CUDA Driver context)
54 that will be registered internally by the driver as an additional device (child)
55 connected to a physical device (parent).
56 Each device (parent or child) is represented through a ID
57 required to indicate which device a given operation should be executed on.
62 gpudev can allocate on an input given GPU device a memory area
63 returning the pointer to that memory.
64 Later, it's also possible to free that memory with gpudev.
65 GPU memory allocated outside of the gpudev library
66 (e.g. with GPU-specific library) cannot be freed by the gpudev library.
71 gpudev can register a CPU memory area to make it visible from a GPU device.
72 Later, it's also possible to unregister that memory with gpudev.
73 CPU memory registered outside of the gpudev library
74 (e.g. with GPU specific library) cannot be unregistered by the gpudev library.
79 gpudev can map into the CPU address space a GPU memory address allocated with gpudev.
80 gpudev returns a pointer the CPU can use to access (ready or write) GPU memory.
81 Later, it's also possible to unmap that memory with gpudev.
82 GPU memory CPU mapped outside of the gpudev library (e.g. with GPU specific library)
83 cannot be unmapped by the gpudev library.
88 Some GPU drivers may need, under certain conditions,
89 to enforce the coherency of external devices writes (e.g. NIC receiving packets)
91 gpudev abstracts and exposes this capability.
96 Considering an application with some GPU task
97 that's waiting to receive a signal from the CPU
98 to move forward with the execution.
99 The communication flag allocates a CPU memory GPU-visible ``uint32_t`` flag
100 that can be used by the CPU to communicate with a GPU task.
105 By default, DPDK pulls free mbufs from a mempool to receive packets.
106 Best practice, especially in a multithreaded application,
107 is to no make any assumption on which mbufs will be used
108 to receive the next bursts of packets.
109 Considering an application with a GPU memory mempool
110 attached to a receive queue having some task waiting on the GPU
111 to receive a new burst of packets to be processed,
112 there is the need to communicate from the CPU
113 the list of mbuf payload addresses where received packet have been stored.
114 The ``rte_gpu_comm_*()`` functions are responsible to create a list of packets
115 that can be populated with receive mbuf payload addresses
116 and communicated to the task running on the GPU.
122 In the example below, there is a pseudo-code to give an example
123 about how to use functions in this library in case of a CUDA application.
127 //////////////////////////////////////////////////////////////////////////
128 ///// gpudev library + CUDA functions
129 //////////////////////////////////////////////////////////////////////////
130 #define GPU_PAGE_SHIFT 16
131 #define GPU_PAGE_SIZE (1UL << GPU_PAGE_SHIFT)
135 struct rte_gpu_flag quit_flag;
136 struct rte_gpu_comm_list *comm_list;
138 int comm_list_entry = 0;
139 struct rte_mbuf *rx_mbufs[max_rx_mbufs];
140 cudaStream_t cstream;
141 struct rte_mempool *mpool_payload, *mpool_header;
142 struct rte_pktmbuf_extmem ext_mem;
146 /* Initialize CUDA objects (cstream, context, etc..). */
147 /* Use gpudev library to register a new CUDA context if any. */
149 /* Let's assume the application wants to use the default context of the GPU device 0. */
152 /* Create an external memory mempool using memory allocated on the GPU. */
153 ext_mem.elt_size = mbufs_headroom_size;
154 ext_mem.buf_len = RTE_ALIGN_CEIL(mbufs_num * ext_mem.elt_size, GPU_PAGE_SIZE);
155 ext_mem.buf_iova = RTE_BAD_IOVA;
156 ext_mem.buf_ptr = rte_gpu_mem_alloc(dev_id, ext_mem.buf_len, 0);
157 rte_extmem_register(ext_mem.buf_ptr, ext_mem.buf_len, NULL, ext_mem.buf_iova, GPU_PAGE_SIZE);
158 rte_dev_dma_map(rte_eth_devices[port_id].device,
159 ext_mem.buf_ptr, ext_mem.buf_iova, ext_mem.buf_len);
160 mpool_payload = rte_pktmbuf_pool_create_extbuf("gpu_mempool", mbufs_num,
161 0, 0, ext_mem.elt_size,
162 rte_socket_id(), &ext_mem, 1);
165 * Create CPU - device communication flag.
166 * With this flag, the CPU can tell to the CUDA kernel to exit from the main loop.
168 rte_gpu_comm_create_flag(dev_id, &quit_flag, RTE_GPU_COMM_FLAG_CPU);
169 rte_gpu_comm_set_flag(&quit_flag , 0);
172 * Create CPU - device communication list.
173 * Each entry of this list will be populated by the CPU
174 * with a new set of received mbufs that the CUDA kernel has to process.
176 comm_list = rte_gpu_comm_create_list(dev_id, num_entries);
178 /* A very simple CUDA kernel with just 1 CUDA block and RTE_GPU_COMM_LIST_PKTS_MAX CUDA threads. */
179 cuda_kernel_packet_processing<<<1, RTE_GPU_COMM_LIST_PKTS_MAX, 0, cstream>>>(quit_flag->ptr, comm_list, num_entries, ...);
182 * For simplicity, the CPU here receives only 2 bursts of mbufs.
183 * In a real application, network activity and device processing should overlap.
185 nb_rx = rte_eth_rx_burst(port_id, queue_id, &(rx_mbufs[0]), max_rx_mbufs);
186 rte_gpu_comm_populate_list_pkts(comm_list[0], rx_mbufs, nb_rx);
187 nb_rx = rte_eth_rx_burst(port_id, queue_id, &(rx_mbufs[0]), max_rx_mbufs);
188 rte_gpu_comm_populate_list_pkts(comm_list[1], rx_mbufs, nb_rx);
191 * CPU waits for the completion of the packets' processing on the CUDA kernel
192 * and then it does a cleanup of the received mbufs.
194 while (rte_gpu_comm_cleanup_list(comm_list[0]));
195 while (rte_gpu_comm_cleanup_list(comm_list[1]));
197 /* CPU notifies the CUDA kernel that it has to terminate. */
198 rte_gpu_comm_set_flag(&quit_flag, 1);
200 /* gpudev objects cleanup/destruction */
201 rte_gpu_mem_free(dev_id, ext_mem.buf_len);
206 //////////////////////////////////////////////////////////////////////////
208 //////////////////////////////////////////////////////////////////////////
210 void cuda_kernel(uint32_t * quit_flag_ptr, struct rte_gpu_comm_list *comm_list, int comm_list_entries)
212 int comm_list_index = 0;
213 struct rte_gpu_comm_pkt *pkt_list = NULL;
215 /* Do some pre-processing operations. */
217 /* GPU kernel keeps checking this flag to know if it has to quit or wait for more packets. */
218 while (*quit_flag_ptr == 0) {
219 if (comm_list[comm_list_index]->status_d[0] != RTE_GPU_COMM_LIST_READY)
222 if (threadIdx.x < comm_list[comm_list_index]->num_pkts)
224 /* Each CUDA thread processes a different packet. */
225 packet_processing(comm_list[comm_list_index]->addr, comm_list[comm_list_index]->size, ..);
230 /* Wait for new packets on the next communication list entry. */
231 comm_list_index = (comm_list_index+1) % comm_list_entries;
234 /* Do some post-processing operations. */