xref: /dpdk/drivers/net/virtio/virtio_user/vhost_kernel.c (revision 945acb4a0d644d194f1823084a234f9c286dcf8c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2016 Intel Corporation
3  */
4 
5 #include <sys/types.h>
6 #include <sys/stat.h>
7 #include <fcntl.h>
8 #include <unistd.h>
9 
10 #include <rte_memory.h>
11 #include <rte_eal_memconfig.h>
12 
13 #include "vhost.h"
14 #include "virtio_user_dev.h"
15 #include "vhost_kernel_tap.h"
16 
17 struct vhost_memory_kernel {
18 	uint32_t nregions;
19 	uint32_t padding;
20 	struct vhost_memory_region regions[0];
21 };
22 
23 /* vhost kernel ioctls */
24 #define VHOST_VIRTIO 0xAF
25 #define VHOST_GET_FEATURES _IOR(VHOST_VIRTIO, 0x00, __u64)
26 #define VHOST_SET_FEATURES _IOW(VHOST_VIRTIO, 0x00, __u64)
27 #define VHOST_SET_OWNER _IO(VHOST_VIRTIO, 0x01)
28 #define VHOST_RESET_OWNER _IO(VHOST_VIRTIO, 0x02)
29 #define VHOST_SET_MEM_TABLE _IOW(VHOST_VIRTIO, 0x03, struct vhost_memory_kernel)
30 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
31 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
32 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
33 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
34 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
35 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
36 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
37 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
38 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
39 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
40 
41 static uint64_t max_regions = 64;
42 
43 static void
44 get_vhost_kernel_max_regions(void)
45 {
46 	int fd;
47 	char buf[20] = {'\0'};
48 
49 	fd = open("/sys/module/vhost/parameters/max_mem_regions", O_RDONLY);
50 	if (fd < 0)
51 		return;
52 
53 	if (read(fd, buf, sizeof(buf) - 1) > 0)
54 		max_regions = strtoull(buf, NULL, 10);
55 
56 	close(fd);
57 }
58 
59 static uint64_t vhost_req_user_to_kernel[] = {
60 	[VHOST_USER_SET_OWNER] = VHOST_SET_OWNER,
61 	[VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
62 	[VHOST_USER_SET_FEATURES] = VHOST_SET_FEATURES,
63 	[VHOST_USER_GET_FEATURES] = VHOST_GET_FEATURES,
64 	[VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
65 	[VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
66 	[VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
67 	[VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
68 	[VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
69 	[VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
70 	[VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
71 };
72 
73 /* By default, vhost kernel module allows 64 regions, but DPDK allows
74  * 256 segments. As a relief, below function merges those virtually
75  * adjacent memsegs into one region.
76  */
77 static struct vhost_memory_kernel *
78 prepare_vhost_memory_kernel(void)
79 {
80 	uint32_t i, j, k = 0;
81 	struct rte_memseg *seg;
82 	struct vhost_memory_region *mr;
83 	struct vhost_memory_kernel *vm;
84 
85 	vm = malloc(sizeof(struct vhost_memory_kernel) +
86 		    max_regions *
87 		    sizeof(struct vhost_memory_region));
88 	if (!vm)
89 		return NULL;
90 
91 	for (i = 0; i < RTE_MAX_MEMSEG; ++i) {
92 		seg = &rte_eal_get_configuration()->mem_config->memseg[i];
93 		if (!seg->addr)
94 			break;
95 
96 		int new_region = 1;
97 
98 		for (j = 0; j < k; ++j) {
99 			mr = &vm->regions[j];
100 
101 			if (mr->userspace_addr + mr->memory_size ==
102 			    (uint64_t)(uintptr_t)seg->addr) {
103 				mr->memory_size += seg->len;
104 				new_region = 0;
105 				break;
106 			}
107 
108 			if ((uint64_t)(uintptr_t)seg->addr + seg->len ==
109 			    mr->userspace_addr) {
110 				mr->guest_phys_addr =
111 					(uint64_t)(uintptr_t)seg->addr;
112 				mr->userspace_addr =
113 					(uint64_t)(uintptr_t)seg->addr;
114 				mr->memory_size += seg->len;
115 				new_region = 0;
116 				break;
117 			}
118 		}
119 
120 		if (new_region == 0)
121 			continue;
122 
123 		mr = &vm->regions[k++];
124 		/* use vaddr here! */
125 		mr->guest_phys_addr = (uint64_t)(uintptr_t)seg->addr;
126 		mr->userspace_addr = (uint64_t)(uintptr_t)seg->addr;
127 		mr->memory_size = seg->len;
128 		mr->mmap_offset = 0;
129 
130 		if (k >= max_regions) {
131 			free(vm);
132 			return NULL;
133 		}
134 	}
135 
136 	vm->nregions = k;
137 	vm->padding = 0;
138 	return vm;
139 }
140 
141 /* with below features, vhost kernel does not need to do the checksum and TSO,
142  * these info will be passed to virtio_user through virtio net header.
143  */
144 #define VHOST_KERNEL_GUEST_OFFLOADS_MASK	\
145 	((1ULL << VIRTIO_NET_F_GUEST_CSUM) |	\
146 	 (1ULL << VIRTIO_NET_F_GUEST_TSO4) |	\
147 	 (1ULL << VIRTIO_NET_F_GUEST_TSO6) |	\
148 	 (1ULL << VIRTIO_NET_F_GUEST_ECN)  |	\
149 	 (1ULL << VIRTIO_NET_F_GUEST_UFO))
150 
151 /* with below features, when flows from virtio_user to vhost kernel
152  * (1) if flows goes up through the kernel networking stack, it does not need
153  * to verify checksum, which can save CPU cycles;
154  * (2) if flows goes through a Linux bridge and outside from an interface
155  * (kernel driver), checksum and TSO will be done by GSO in kernel or even
156  * offloaded into real physical device.
157  */
158 #define VHOST_KERNEL_HOST_OFFLOADS_MASK		\
159 	((1ULL << VIRTIO_NET_F_HOST_TSO4) |	\
160 	 (1ULL << VIRTIO_NET_F_HOST_TSO6) |	\
161 	 (1ULL << VIRTIO_NET_F_CSUM))
162 
163 static int
164 tap_supporte_mq(void)
165 {
166 	int tapfd;
167 	unsigned int tap_features;
168 
169 	tapfd = open(PATH_NET_TUN, O_RDWR);
170 	if (tapfd < 0) {
171 		PMD_DRV_LOG(ERR, "fail to open %s: %s",
172 			    PATH_NET_TUN, strerror(errno));
173 		return -1;
174 	}
175 
176 	if (ioctl(tapfd, TUNGETFEATURES, &tap_features) == -1) {
177 		PMD_DRV_LOG(ERR, "TUNGETFEATURES failed: %s", strerror(errno));
178 		close(tapfd);
179 		return -1;
180 	}
181 
182 	close(tapfd);
183 	return tap_features & IFF_MULTI_QUEUE;
184 }
185 
186 static int
187 vhost_kernel_ioctl(struct virtio_user_dev *dev,
188 		   enum vhost_user_request req,
189 		   void *arg)
190 {
191 	int ret = -1;
192 	unsigned int i;
193 	uint64_t req_kernel;
194 	struct vhost_memory_kernel *vm = NULL;
195 	int vhostfd;
196 	unsigned int queue_sel;
197 
198 	PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
199 
200 	req_kernel = vhost_req_user_to_kernel[req];
201 
202 	if (req_kernel == VHOST_SET_MEM_TABLE) {
203 		vm = prepare_vhost_memory_kernel();
204 		if (!vm)
205 			return -1;
206 		arg = (void *)vm;
207 	}
208 
209 	if (req_kernel == VHOST_SET_FEATURES) {
210 		/* We don't need memory protection here */
211 		*(uint64_t *)arg &= ~(1ULL << VIRTIO_F_IOMMU_PLATFORM);
212 
213 		/* VHOST kernel does not know about below flags */
214 		*(uint64_t *)arg &= ~VHOST_KERNEL_GUEST_OFFLOADS_MASK;
215 		*(uint64_t *)arg &= ~VHOST_KERNEL_HOST_OFFLOADS_MASK;
216 
217 		*(uint64_t *)arg &= ~(1ULL << VIRTIO_NET_F_MQ);
218 	}
219 
220 	switch (req_kernel) {
221 	case VHOST_SET_VRING_NUM:
222 	case VHOST_SET_VRING_ADDR:
223 	case VHOST_SET_VRING_BASE:
224 	case VHOST_GET_VRING_BASE:
225 	case VHOST_SET_VRING_KICK:
226 	case VHOST_SET_VRING_CALL:
227 		queue_sel = *(unsigned int *)arg;
228 		vhostfd = dev->vhostfds[queue_sel / 2];
229 		*(unsigned int *)arg = queue_sel % 2;
230 		PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
231 			    vhostfd, *(unsigned int *)arg);
232 		break;
233 	default:
234 		vhostfd = -1;
235 	}
236 	if (vhostfd == -1) {
237 		for (i = 0; i < dev->max_queue_pairs; ++i) {
238 			if (dev->vhostfds[i] < 0)
239 				continue;
240 
241 			ret = ioctl(dev->vhostfds[i], req_kernel, arg);
242 			if (ret < 0)
243 				break;
244 		}
245 	} else {
246 		ret = ioctl(vhostfd, req_kernel, arg);
247 	}
248 
249 	if (!ret && req_kernel == VHOST_GET_FEATURES) {
250 		/* with tap as the backend, all these features are supported
251 		 * but not claimed by vhost-net, so we add them back when
252 		 * reporting to upper layer.
253 		 */
254 		*((uint64_t *)arg) |= VHOST_KERNEL_GUEST_OFFLOADS_MASK;
255 		*((uint64_t *)arg) |= VHOST_KERNEL_HOST_OFFLOADS_MASK;
256 
257 		/* vhost_kernel will not declare this feature, but it does
258 		 * support multi-queue.
259 		 */
260 		if (tap_supporte_mq())
261 			*(uint64_t *)arg |= (1ull << VIRTIO_NET_F_MQ);
262 	}
263 
264 	if (vm)
265 		free(vm);
266 
267 	if (ret < 0)
268 		PMD_DRV_LOG(ERR, "%s failed: %s",
269 			    vhost_msg_strings[req], strerror(errno));
270 
271 	return ret;
272 }
273 
274 /**
275  * Set up environment to talk with a vhost kernel backend.
276  *
277  * @return
278  *   - (-1) if fail to set up;
279  *   - (>=0) if successful.
280  */
281 static int
282 vhost_kernel_setup(struct virtio_user_dev *dev)
283 {
284 	int vhostfd;
285 	uint32_t i;
286 
287 	get_vhost_kernel_max_regions();
288 
289 	for (i = 0; i < dev->max_queue_pairs; ++i) {
290 		vhostfd = open(dev->path, O_RDWR);
291 		if (vhostfd < 0) {
292 			PMD_DRV_LOG(ERR, "fail to open %s, %s",
293 				    dev->path, strerror(errno));
294 			return -1;
295 		}
296 
297 		dev->vhostfds[i] = vhostfd;
298 	}
299 
300 	return 0;
301 }
302 
303 static int
304 vhost_kernel_set_backend(int vhostfd, int tapfd)
305 {
306 	struct vhost_vring_file f;
307 
308 	f.fd = tapfd;
309 	f.index = 0;
310 	if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
311 		PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
312 				strerror(errno));
313 		return -1;
314 	}
315 
316 	f.index = 1;
317 	if (ioctl(vhostfd, VHOST_NET_SET_BACKEND, &f) < 0) {
318 		PMD_DRV_LOG(ERR, "VHOST_NET_SET_BACKEND fails, %s",
319 				strerror(errno));
320 		return -1;
321 	}
322 
323 	return 0;
324 }
325 
326 static int
327 vhost_kernel_enable_queue_pair(struct virtio_user_dev *dev,
328 			       uint16_t pair_idx,
329 			       int enable)
330 {
331 	int hdr_size;
332 	int vhostfd;
333 	int tapfd;
334 	int req_mq = (dev->max_queue_pairs > 1);
335 
336 	vhostfd = dev->vhostfds[pair_idx];
337 
338 	if (!enable) {
339 		if (dev->tapfds[pair_idx] >= 0) {
340 			close(dev->tapfds[pair_idx]);
341 			dev->tapfds[pair_idx] = -1;
342 		}
343 		return vhost_kernel_set_backend(vhostfd, -1);
344 	} else if (dev->tapfds[pair_idx] >= 0) {
345 		return 0;
346 	}
347 
348 	if ((dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF)) ||
349 	    (dev->features & (1ULL << VIRTIO_F_VERSION_1)))
350 		hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
351 	else
352 		hdr_size = sizeof(struct virtio_net_hdr);
353 
354 	tapfd = vhost_kernel_open_tap(&dev->ifname, hdr_size, req_mq);
355 	if (tapfd < 0) {
356 		PMD_DRV_LOG(ERR, "fail to open tap for vhost kernel");
357 		return -1;
358 	}
359 
360 	if (vhost_kernel_set_backend(vhostfd, tapfd) < 0) {
361 		PMD_DRV_LOG(ERR, "fail to set backend for vhost kernel");
362 		close(tapfd);
363 		return -1;
364 	}
365 
366 	dev->tapfds[pair_idx] = tapfd;
367 	return 0;
368 }
369 
370 struct virtio_user_backend_ops ops_kernel = {
371 	.setup = vhost_kernel_setup,
372 	.send_request = vhost_kernel_ioctl,
373 	.enable_qp = vhost_kernel_enable_queue_pair
374 };
375