1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2020 Red Hat Inc.
3  */
4 
5 #include <sys/ioctl.h>
6 #include <sys/types.h>
7 #include <sys/stat.h>
8 #include <fcntl.h>
9 #include <unistd.h>
10 
11 #include <rte_memory.h>
12 
13 #include "vhost.h"
14 #include "virtio_user_dev.h"
15 
16 /* vhost kernel & vdpa ioctls */
17 #define VHOST_VIRTIO 0xAF
18 #define VHOST_GET_FEATURES _IOR(VHOST_VIRTIO, 0x00, __u64)
19 #define VHOST_SET_FEATURES _IOW(VHOST_VIRTIO, 0x00, __u64)
20 #define VHOST_SET_OWNER _IO(VHOST_VIRTIO, 0x01)
21 #define VHOST_RESET_OWNER _IO(VHOST_VIRTIO, 0x02)
22 #define VHOST_SET_MEM_TABLE _IOW(VHOST_VIRTIO, 0x03, void *)
23 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
24 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
25 #define VHOST_SET_VRING_NUM _IOW(VHOST_VIRTIO, 0x10, struct vhost_vring_state)
26 #define VHOST_SET_VRING_ADDR _IOW(VHOST_VIRTIO, 0x11, struct vhost_vring_addr)
27 #define VHOST_SET_VRING_BASE _IOW(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
28 #define VHOST_GET_VRING_BASE _IOWR(VHOST_VIRTIO, 0x12, struct vhost_vring_state)
29 #define VHOST_SET_VRING_KICK _IOW(VHOST_VIRTIO, 0x20, struct vhost_vring_file)
30 #define VHOST_SET_VRING_CALL _IOW(VHOST_VIRTIO, 0x21, struct vhost_vring_file)
31 #define VHOST_SET_VRING_ERR _IOW(VHOST_VIRTIO, 0x22, struct vhost_vring_file)
32 #define VHOST_NET_SET_BACKEND _IOW(VHOST_VIRTIO, 0x30, struct vhost_vring_file)
33 #define VHOST_VDPA_GET_DEVICE_ID _IOR(VHOST_VIRTIO, 0x70, __u32)
34 #define VHOST_VDPA_GET_STATUS _IOR(VHOST_VIRTIO, 0x71, __u8)
35 #define VHOST_VDPA_SET_STATUS _IOW(VHOST_VIRTIO, 0x72, __u8)
36 #define VHOST_VDPA_SET_VRING_ENABLE	_IOW(VHOST_VIRTIO, 0x75, \
37 					     struct vhost_vring_state)
38 
39 static uint64_t vhost_req_user_to_vdpa[] = {
40 	[VHOST_USER_SET_OWNER] = VHOST_SET_OWNER,
41 	[VHOST_USER_RESET_OWNER] = VHOST_RESET_OWNER,
42 	[VHOST_USER_SET_FEATURES] = VHOST_SET_FEATURES,
43 	[VHOST_USER_GET_FEATURES] = VHOST_GET_FEATURES,
44 	[VHOST_USER_SET_VRING_CALL] = VHOST_SET_VRING_CALL,
45 	[VHOST_USER_SET_VRING_NUM] = VHOST_SET_VRING_NUM,
46 	[VHOST_USER_SET_VRING_BASE] = VHOST_SET_VRING_BASE,
47 	[VHOST_USER_GET_VRING_BASE] = VHOST_GET_VRING_BASE,
48 	[VHOST_USER_SET_VRING_ADDR] = VHOST_SET_VRING_ADDR,
49 	[VHOST_USER_SET_VRING_KICK] = VHOST_SET_VRING_KICK,
50 	[VHOST_USER_SET_MEM_TABLE] = VHOST_SET_MEM_TABLE,
51 	[VHOST_USER_SET_STATUS] = VHOST_VDPA_SET_STATUS,
52 	[VHOST_USER_GET_STATUS] = VHOST_VDPA_GET_STATUS,
53 	[VHOST_USER_SET_VRING_ENABLE] = VHOST_VDPA_SET_VRING_ENABLE,
54 };
55 
56 /* no alignment requirement */
57 struct vhost_iotlb_msg {
58 	uint64_t iova;
59 	uint64_t size;
60 	uint64_t uaddr;
61 #define VHOST_ACCESS_RO      0x1
62 #define VHOST_ACCESS_WO      0x2
63 #define VHOST_ACCESS_RW      0x3
64 	uint8_t perm;
65 #define VHOST_IOTLB_MISS           1
66 #define VHOST_IOTLB_UPDATE         2
67 #define VHOST_IOTLB_INVALIDATE     3
68 #define VHOST_IOTLB_ACCESS_FAIL    4
69 	uint8_t type;
70 };
71 
72 #define VHOST_IOTLB_MSG_V2 0x2
73 
74 struct vhost_msg {
75 	uint32_t type;
76 	uint32_t reserved;
77 	union {
78 		struct vhost_iotlb_msg iotlb;
79 		uint8_t padding[64];
80 	};
81 };
82 
83 static int
vhost_vdpa_dma_map(struct virtio_user_dev * dev,void * addr,uint64_t iova,size_t len)84 vhost_vdpa_dma_map(struct virtio_user_dev *dev, void *addr,
85 				  uint64_t iova, size_t len)
86 {
87 	struct vhost_msg msg = {};
88 
89 	msg.type = VHOST_IOTLB_MSG_V2;
90 	msg.iotlb.type = VHOST_IOTLB_UPDATE;
91 	msg.iotlb.iova = iova;
92 	msg.iotlb.uaddr = (uint64_t)(uintptr_t)addr;
93 	msg.iotlb.size = len;
94 	msg.iotlb.perm = VHOST_ACCESS_RW;
95 
96 	if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
97 		PMD_DRV_LOG(ERR, "Failed to send IOTLB update (%s)",
98 				strerror(errno));
99 		return -1;
100 	}
101 
102 	return 0;
103 }
104 
105 static int
vhost_vdpa_dma_unmap(struct virtio_user_dev * dev,__rte_unused void * addr,uint64_t iova,size_t len)106 vhost_vdpa_dma_unmap(struct virtio_user_dev *dev, __rte_unused void *addr,
107 				  uint64_t iova, size_t len)
108 {
109 	struct vhost_msg msg = {};
110 
111 	msg.type = VHOST_IOTLB_MSG_V2;
112 	msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
113 	msg.iotlb.iova = iova;
114 	msg.iotlb.size = len;
115 
116 	if (write(dev->vhostfd, &msg, sizeof(msg)) != sizeof(msg)) {
117 		PMD_DRV_LOG(ERR, "Failed to send IOTLB invalidate (%s)",
118 				strerror(errno));
119 		return -1;
120 	}
121 
122 	return 0;
123 }
124 
125 
126 static int
vhost_vdpa_map_contig(const struct rte_memseg_list * msl,const struct rte_memseg * ms,size_t len,void * arg)127 vhost_vdpa_map_contig(const struct rte_memseg_list *msl,
128 		const struct rte_memseg *ms, size_t len, void *arg)
129 {
130 	struct virtio_user_dev *dev = arg;
131 
132 	if (msl->external)
133 		return 0;
134 
135 	return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, len);
136 }
137 
138 static int
vhost_vdpa_map(const struct rte_memseg_list * msl,const struct rte_memseg * ms,void * arg)139 vhost_vdpa_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
140 		void *arg)
141 {
142 	struct virtio_user_dev *dev = arg;
143 
144 	/* skip external memory that isn't a heap */
145 	if (msl->external && !msl->heap)
146 		return 0;
147 
148 	/* skip any segments with invalid IOVA addresses */
149 	if (ms->iova == RTE_BAD_IOVA)
150 		return 0;
151 
152 	/* if IOVA mode is VA, we've already mapped the internal segments */
153 	if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
154 		return 0;
155 
156 	return vhost_vdpa_dma_map(dev, ms->addr, ms->iova, ms->len);
157 }
158 
159 static int
vhost_vdpa_dma_map_all(struct virtio_user_dev * dev)160 vhost_vdpa_dma_map_all(struct virtio_user_dev *dev)
161 {
162 	vhost_vdpa_dma_unmap(dev, NULL, 0, SIZE_MAX);
163 
164 	if (rte_eal_iova_mode() == RTE_IOVA_VA) {
165 		/* with IOVA as VA mode, we can get away with mapping contiguous
166 		 * chunks rather than going page-by-page.
167 		 */
168 		int ret = rte_memseg_contig_walk_thread_unsafe(
169 				vhost_vdpa_map_contig, dev);
170 		if (ret)
171 			return ret;
172 		/* we have to continue the walk because we've skipped the
173 		 * external segments during the config walk.
174 		 */
175 	}
176 	return rte_memseg_walk_thread_unsafe(vhost_vdpa_map, dev);
177 }
178 
179 /* with below features, vhost vdpa does not need to do the checksum and TSO,
180  * these info will be passed to virtio_user through virtio net header.
181  */
182 #define VHOST_VDPA_GUEST_OFFLOADS_MASK	\
183 	((1ULL << VIRTIO_NET_F_GUEST_CSUM) |	\
184 	 (1ULL << VIRTIO_NET_F_GUEST_TSO4) |	\
185 	 (1ULL << VIRTIO_NET_F_GUEST_TSO6) |	\
186 	 (1ULL << VIRTIO_NET_F_GUEST_ECN)  |	\
187 	 (1ULL << VIRTIO_NET_F_GUEST_UFO))
188 
189 #define VHOST_VDPA_HOST_OFFLOADS_MASK		\
190 	((1ULL << VIRTIO_NET_F_HOST_TSO4) |	\
191 	 (1ULL << VIRTIO_NET_F_HOST_TSO6) |	\
192 	 (1ULL << VIRTIO_NET_F_CSUM))
193 
194 static int
vhost_vdpa_ioctl(struct virtio_user_dev * dev,enum vhost_user_request req,void * arg)195 vhost_vdpa_ioctl(struct virtio_user_dev *dev,
196 		   enum vhost_user_request req,
197 		   void *arg)
198 {
199 	int ret = -1;
200 	uint64_t req_vdpa;
201 
202 	PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
203 
204 	req_vdpa = vhost_req_user_to_vdpa[req];
205 
206 	if (req_vdpa == VHOST_SET_MEM_TABLE)
207 		return vhost_vdpa_dma_map_all(dev);
208 
209 	if (req_vdpa == VHOST_SET_FEATURES) {
210 		/* WORKAROUND */
211 		*(uint64_t *)arg |= 1ULL << VIRTIO_F_IOMMU_PLATFORM;
212 
213 		/* Multiqueue not supported for now */
214 		*(uint64_t *)arg &= ~(1ULL << VIRTIO_NET_F_MQ);
215 	}
216 
217 	switch (req_vdpa) {
218 	case VHOST_SET_VRING_NUM:
219 	case VHOST_SET_VRING_ADDR:
220 	case VHOST_SET_VRING_BASE:
221 	case VHOST_GET_VRING_BASE:
222 	case VHOST_SET_VRING_KICK:
223 	case VHOST_SET_VRING_CALL:
224 		PMD_DRV_LOG(DEBUG, "vhostfd=%d, index=%u",
225 			    dev->vhostfd, *(unsigned int *)arg);
226 		break;
227 	default:
228 		break;
229 	}
230 
231 	ret = ioctl(dev->vhostfd, req_vdpa, arg);
232 	if (ret < 0)
233 		PMD_DRV_LOG(ERR, "%s failed: %s",
234 			    vhost_msg_strings[req], strerror(errno));
235 
236 	return ret;
237 }
238 
239 /**
240  * Set up environment to talk with a vhost vdpa backend.
241  *
242  * @return
243  *   - (-1) if fail to set up;
244  *   - (>=0) if successful.
245  */
246 static int
vhost_vdpa_setup(struct virtio_user_dev * dev)247 vhost_vdpa_setup(struct virtio_user_dev *dev)
248 {
249 	uint32_t did = (uint32_t)-1;
250 
251 	dev->vhostfd = open(dev->path, O_RDWR);
252 	if (dev->vhostfd < 0) {
253 		PMD_DRV_LOG(ERR, "Failed to open %s: %s\n",
254 				dev->path, strerror(errno));
255 		return -1;
256 	}
257 
258 	if (ioctl(dev->vhostfd, VHOST_VDPA_GET_DEVICE_ID, &did) < 0 ||
259 			did != VIRTIO_ID_NETWORK) {
260 		PMD_DRV_LOG(ERR, "Invalid vdpa device ID: %u\n", did);
261 		return -1;
262 	}
263 
264 	return 0;
265 }
266 
267 static int
vhost_vdpa_enable_queue_pair(struct virtio_user_dev * dev,uint16_t pair_idx,int enable)268 vhost_vdpa_enable_queue_pair(struct virtio_user_dev *dev,
269 			       uint16_t pair_idx,
270 			       int enable)
271 {
272 	int i;
273 
274 	if (dev->qp_enabled[pair_idx] == enable)
275 		return 0;
276 
277 	for (i = 0; i < 2; ++i) {
278 		struct vhost_vring_state state = {
279 			.index = pair_idx * 2 + i,
280 			.num   = enable,
281 		};
282 
283 		if (vhost_vdpa_ioctl(dev, VHOST_USER_SET_VRING_ENABLE, &state))
284 			return -1;
285 	}
286 
287 	dev->qp_enabled[pair_idx] = enable;
288 
289 	return 0;
290 }
291 
292 struct virtio_user_backend_ops virtio_ops_vdpa = {
293 	.setup = vhost_vdpa_setup,
294 	.send_request = vhost_vdpa_ioctl,
295 	.enable_qp = vhost_vdpa_enable_queue_pair,
296 	.dma_map = vhost_vdpa_dma_map,
297 	.dma_unmap = vhost_vdpa_dma_unmap,
298 };
299