1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2017 Intel Corporation 3 */ 4 5 #include <stdint.h> 6 #include <stdbool.h> 7 #include <linux/virtio_net.h> 8 9 #include <rte_mbuf.h> 10 #include <rte_memcpy.h> 11 #include <rte_vhost.h> 12 13 #include "main.h" 14 15 /* 16 * A very simple vhost-user net driver implementation, without 17 * any extra features being enabled, such as TSO and mrg-Rx. 18 */ 19 20 void 21 vs_vhost_net_setup(struct vhost_dev *dev) 22 { 23 uint16_t i; 24 int vid = dev->vid; 25 struct vhost_queue *queue; 26 27 RTE_LOG(INFO, VHOST_CONFIG, 28 "setting builtin vhost-user net driver\n"); 29 30 rte_vhost_get_negotiated_features(vid, &dev->features); 31 if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) 32 dev->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 33 else 34 dev->hdr_len = sizeof(struct virtio_net_hdr); 35 36 rte_vhost_get_mem_table(vid, &dev->mem); 37 38 dev->nr_vrings = rte_vhost_get_vring_num(vid); 39 for (i = 0; i < dev->nr_vrings; i++) { 40 queue = &dev->queues[i]; 41 42 queue->last_used_idx = 0; 43 queue->last_avail_idx = 0; 44 rte_vhost_get_vhost_vring(vid, i, &queue->vr); 45 } 46 } 47 48 void 49 vs_vhost_net_remove(struct vhost_dev *dev) 50 { 51 free(dev->mem); 52 } 53 54 static __rte_always_inline int 55 enqueue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr, 56 struct rte_mbuf *m, uint16_t desc_idx) 57 { 58 uint32_t desc_avail, desc_offset; 59 uint64_t desc_chunck_len; 60 uint32_t mbuf_avail, mbuf_offset; 61 uint32_t cpy_len; 62 struct vring_desc *desc; 63 uint64_t desc_addr, desc_gaddr; 64 struct virtio_net_hdr virtio_hdr = {0, 0, 0, 0, 0, 0}; 65 /* A counter to avoid desc dead loop chain */ 66 uint16_t nr_desc = 1; 67 68 desc = &vr->desc[desc_idx]; 69 desc_chunck_len = desc->len; 70 desc_gaddr = desc->addr; 71 desc_addr = rte_vhost_va_from_guest_pa( 72 dev->mem, desc_gaddr, &desc_chunck_len); 73 /* 74 * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid 75 * performance issue with some versions of gcc (4.8.4 and 5.3.0) which 76 * otherwise stores offset on the stack instead of in a register. 77 */ 78 if (unlikely(desc->len < dev->hdr_len) || !desc_addr) 79 return -1; 80 81 rte_prefetch0((void *)(uintptr_t)desc_addr); 82 83 /* write virtio-net header */ 84 if (likely(desc_chunck_len >= dev->hdr_len)) { 85 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = virtio_hdr; 86 desc_offset = dev->hdr_len; 87 } else { 88 uint64_t len; 89 uint64_t remain = dev->hdr_len; 90 uint64_t src = (uint64_t)(uintptr_t)&virtio_hdr, dst; 91 uint64_t guest_addr = desc_gaddr; 92 93 while (remain) { 94 len = remain; 95 dst = rte_vhost_va_from_guest_pa(dev->mem, 96 guest_addr, &len); 97 if (unlikely(!dst || !len)) 98 return -1; 99 100 rte_memcpy((void *)(uintptr_t)dst, 101 (void *)(uintptr_t)src, 102 len); 103 104 remain -= len; 105 guest_addr += len; 106 src += len; 107 } 108 109 desc_chunck_len = desc->len - dev->hdr_len; 110 desc_gaddr += dev->hdr_len; 111 desc_addr = rte_vhost_va_from_guest_pa( 112 dev->mem, desc_gaddr, 113 &desc_chunck_len); 114 if (unlikely(!desc_addr)) 115 return -1; 116 117 desc_offset = 0; 118 } 119 120 desc_avail = desc->len - dev->hdr_len; 121 122 mbuf_avail = rte_pktmbuf_data_len(m); 123 mbuf_offset = 0; 124 while (mbuf_avail != 0 || m->next != NULL) { 125 /* done with current mbuf, fetch next */ 126 if (mbuf_avail == 0) { 127 m = m->next; 128 129 mbuf_offset = 0; 130 mbuf_avail = rte_pktmbuf_data_len(m); 131 } 132 133 /* done with current desc buf, fetch next */ 134 if (desc_avail == 0) { 135 if ((desc->flags & VRING_DESC_F_NEXT) == 0) { 136 /* Room in vring buffer is not enough */ 137 return -1; 138 } 139 if (unlikely(desc->next >= vr->size || 140 ++nr_desc > vr->size)) 141 return -1; 142 143 desc = &vr->desc[desc->next]; 144 desc_chunck_len = desc->len; 145 desc_gaddr = desc->addr; 146 desc_addr = rte_vhost_va_from_guest_pa( 147 dev->mem, desc_gaddr, &desc_chunck_len); 148 if (unlikely(!desc_addr)) 149 return -1; 150 151 desc_offset = 0; 152 desc_avail = desc->len; 153 } else if (unlikely(desc_chunck_len == 0)) { 154 desc_chunck_len = desc_avail; 155 desc_gaddr += desc_offset; 156 desc_addr = rte_vhost_va_from_guest_pa(dev->mem, 157 desc_gaddr, 158 &desc_chunck_len); 159 if (unlikely(!desc_addr)) 160 return -1; 161 162 desc_offset = 0; 163 } 164 165 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail); 166 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)), 167 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset), 168 cpy_len); 169 170 mbuf_avail -= cpy_len; 171 mbuf_offset += cpy_len; 172 desc_avail -= cpy_len; 173 desc_offset += cpy_len; 174 desc_chunck_len -= cpy_len; 175 } 176 177 return 0; 178 } 179 180 uint16_t 181 vs_enqueue_pkts(struct vhost_dev *dev, uint16_t queue_id, 182 struct rte_mbuf **pkts, uint32_t count) 183 { 184 struct vhost_queue *queue; 185 struct rte_vhost_vring *vr; 186 uint16_t avail_idx, free_entries, start_idx; 187 uint16_t desc_indexes[MAX_PKT_BURST]; 188 uint16_t used_idx; 189 uint32_t i; 190 191 queue = &dev->queues[queue_id]; 192 vr = &queue->vr; 193 194 avail_idx = __atomic_load_n(&vr->avail->idx, __ATOMIC_ACQUIRE); 195 start_idx = queue->last_used_idx; 196 free_entries = avail_idx - start_idx; 197 count = RTE_MIN(count, free_entries); 198 count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST); 199 if (count == 0) 200 return 0; 201 202 /* Retrieve all of the desc indexes first to avoid caching issues. */ 203 rte_prefetch0(&vr->avail->ring[start_idx & (vr->size - 1)]); 204 for (i = 0; i < count; i++) { 205 used_idx = (start_idx + i) & (vr->size - 1); 206 desc_indexes[i] = vr->avail->ring[used_idx]; 207 vr->used->ring[used_idx].id = desc_indexes[i]; 208 vr->used->ring[used_idx].len = pkts[i]->pkt_len + 209 dev->hdr_len; 210 } 211 212 rte_prefetch0(&vr->desc[desc_indexes[0]]); 213 for (i = 0; i < count; i++) { 214 uint16_t desc_idx = desc_indexes[i]; 215 int err; 216 217 err = enqueue_pkt(dev, vr, pkts[i], desc_idx); 218 if (unlikely(err)) { 219 used_idx = (start_idx + i) & (vr->size - 1); 220 vr->used->ring[used_idx].len = dev->hdr_len; 221 } 222 223 if (i + 1 < count) 224 rte_prefetch0(&vr->desc[desc_indexes[i+1]]); 225 } 226 227 __atomic_add_fetch(&vr->used->idx, count, __ATOMIC_RELEASE); 228 queue->last_used_idx += count; 229 230 rte_vhost_vring_call(dev->vid, queue_id); 231 232 return count; 233 } 234 235 static __rte_always_inline int 236 dequeue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr, 237 struct rte_mbuf *m, uint16_t desc_idx, 238 struct rte_mempool *mbuf_pool) 239 { 240 struct vring_desc *desc; 241 uint64_t desc_addr, desc_gaddr; 242 uint32_t desc_avail, desc_offset; 243 uint64_t desc_chunck_len; 244 uint32_t mbuf_avail, mbuf_offset; 245 uint32_t cpy_len; 246 struct rte_mbuf *cur = m, *prev = m; 247 /* A counter to avoid desc dead loop chain */ 248 uint32_t nr_desc = 1; 249 250 desc = &vr->desc[desc_idx]; 251 if (unlikely((desc->len < dev->hdr_len)) || 252 (desc->flags & VRING_DESC_F_INDIRECT)) 253 return -1; 254 255 desc_chunck_len = desc->len; 256 desc_gaddr = desc->addr; 257 desc_addr = rte_vhost_va_from_guest_pa( 258 dev->mem, desc_gaddr, &desc_chunck_len); 259 if (unlikely(!desc_addr)) 260 return -1; 261 262 /* 263 * We don't support ANY_LAYOUT, neither VERSION_1, meaning 264 * a Tx packet from guest must have 2 desc buffers at least: 265 * the first for storing the header and the others for 266 * storing the data. 267 * 268 * And since we don't support TSO, we could simply skip the 269 * header. 270 */ 271 desc = &vr->desc[desc->next]; 272 desc_chunck_len = desc->len; 273 desc_gaddr = desc->addr; 274 desc_addr = rte_vhost_va_from_guest_pa( 275 dev->mem, desc_gaddr, &desc_chunck_len); 276 if (unlikely(!desc_addr)) 277 return -1; 278 rte_prefetch0((void *)(uintptr_t)desc_addr); 279 280 desc_offset = 0; 281 desc_avail = desc->len; 282 nr_desc += 1; 283 284 mbuf_offset = 0; 285 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM; 286 while (1) { 287 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail); 288 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, 289 mbuf_offset), 290 (void *)((uintptr_t)(desc_addr + desc_offset)), 291 cpy_len); 292 293 mbuf_avail -= cpy_len; 294 mbuf_offset += cpy_len; 295 desc_avail -= cpy_len; 296 desc_offset += cpy_len; 297 desc_chunck_len -= cpy_len; 298 299 /* This desc reaches to its end, get the next one */ 300 if (desc_avail == 0) { 301 if ((desc->flags & VRING_DESC_F_NEXT) == 0) 302 break; 303 304 if (unlikely(desc->next >= vr->size || 305 ++nr_desc > vr->size)) 306 return -1; 307 desc = &vr->desc[desc->next]; 308 309 desc_chunck_len = desc->len; 310 desc_gaddr = desc->addr; 311 desc_addr = rte_vhost_va_from_guest_pa( 312 dev->mem, desc_gaddr, &desc_chunck_len); 313 if (unlikely(!desc_addr)) 314 return -1; 315 rte_prefetch0((void *)(uintptr_t)desc_addr); 316 317 desc_offset = 0; 318 desc_avail = desc->len; 319 } else if (unlikely(desc_chunck_len == 0)) { 320 desc_chunck_len = desc_avail; 321 desc_gaddr += desc_offset; 322 desc_addr = rte_vhost_va_from_guest_pa(dev->mem, 323 desc_gaddr, 324 &desc_chunck_len); 325 if (unlikely(!desc_addr)) 326 return -1; 327 328 desc_offset = 0; 329 } 330 331 /* 332 * This mbuf reaches to its end, get a new one 333 * to hold more data. 334 */ 335 if (mbuf_avail == 0) { 336 cur = rte_pktmbuf_alloc(mbuf_pool); 337 if (unlikely(cur == NULL)) { 338 RTE_LOG(ERR, VHOST_DATA, "Failed to " 339 "allocate memory for mbuf.\n"); 340 return -1; 341 } 342 343 prev->next = cur; 344 prev->data_len = mbuf_offset; 345 m->nb_segs += 1; 346 m->pkt_len += mbuf_offset; 347 prev = cur; 348 349 mbuf_offset = 0; 350 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM; 351 } 352 } 353 354 prev->data_len = mbuf_offset; 355 m->pkt_len += mbuf_offset; 356 357 return 0; 358 } 359 360 uint16_t 361 vs_dequeue_pkts(struct vhost_dev *dev, uint16_t queue_id, 362 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count) 363 { 364 struct vhost_queue *queue; 365 struct rte_vhost_vring *vr; 366 uint32_t desc_indexes[MAX_PKT_BURST]; 367 uint32_t used_idx; 368 uint32_t i = 0; 369 uint16_t free_entries; 370 uint16_t avail_idx; 371 372 queue = &dev->queues[queue_id]; 373 vr = &queue->vr; 374 375 free_entries = __atomic_load_n(&vr->avail->idx, __ATOMIC_ACQUIRE) - 376 queue->last_avail_idx; 377 if (free_entries == 0) 378 return 0; 379 380 /* Prefetch available and used ring */ 381 avail_idx = queue->last_avail_idx & (vr->size - 1); 382 used_idx = queue->last_used_idx & (vr->size - 1); 383 rte_prefetch0(&vr->avail->ring[avail_idx]); 384 rte_prefetch0(&vr->used->ring[used_idx]); 385 386 count = RTE_MIN(count, MAX_PKT_BURST); 387 count = RTE_MIN(count, free_entries); 388 389 if (unlikely(count == 0)) 390 return 0; 391 392 /* 393 * Retrieve all of the head indexes first and pre-update used entries 394 * to avoid caching issues. 395 */ 396 for (i = 0; i < count; i++) { 397 avail_idx = (queue->last_avail_idx + i) & (vr->size - 1); 398 used_idx = (queue->last_used_idx + i) & (vr->size - 1); 399 desc_indexes[i] = vr->avail->ring[avail_idx]; 400 401 vr->used->ring[used_idx].id = desc_indexes[i]; 402 vr->used->ring[used_idx].len = 0; 403 } 404 405 /* Prefetch descriptor index. */ 406 rte_prefetch0(&vr->desc[desc_indexes[0]]); 407 for (i = 0; i < count; i++) { 408 int err; 409 410 if (likely(i + 1 < count)) 411 rte_prefetch0(&vr->desc[desc_indexes[i + 1]]); 412 413 pkts[i] = rte_pktmbuf_alloc(mbuf_pool); 414 if (unlikely(pkts[i] == NULL)) { 415 RTE_LOG(ERR, VHOST_DATA, 416 "Failed to allocate memory for mbuf.\n"); 417 break; 418 } 419 420 err = dequeue_pkt(dev, vr, pkts[i], desc_indexes[i], mbuf_pool); 421 if (unlikely(err)) { 422 rte_pktmbuf_free(pkts[i]); 423 break; 424 } 425 426 } 427 428 queue->last_avail_idx += i; 429 queue->last_used_idx += i; 430 431 __atomic_add_fetch(&vr->used->idx, i, __ATOMIC_ACQ_REL); 432 433 rte_vhost_vring_call(dev->vid, queue_id); 434 435 return i; 436 } 437