1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2014 Intel Corporation 3 */ 4 5 #ifndef _VIRTQUEUE_H_ 6 #define _VIRTQUEUE_H_ 7 8 #include <stdint.h> 9 10 #include <rte_atomic.h> 11 #include <rte_memory.h> 12 #include <rte_mempool.h> 13 #include <rte_net.h> 14 15 #include "virtio_pci.h" 16 #include "virtio_ring.h" 17 #include "virtio_logs.h" 18 #include "virtio_rxtx.h" 19 20 struct rte_mbuf; 21 22 #define DEFAULT_TX_FREE_THRESH 32 23 #define DEFAULT_RX_FREE_THRESH 32 24 25 #define VIRTIO_MBUF_BURST_SZ 64 26 /* 27 * Per virtio_ring.h in Linux. 28 * For virtio_pci on SMP, we don't need to order with respect to MMIO 29 * accesses through relaxed memory I/O windows, so smp_mb() et al are 30 * sufficient. 31 * 32 * For using virtio to talk to real devices (eg. vDPA) we do need real 33 * barriers. 34 */ 35 static inline void 36 virtio_mb(uint8_t weak_barriers) 37 { 38 if (weak_barriers) 39 rte_smp_mb(); 40 else 41 rte_mb(); 42 } 43 44 static inline void 45 virtio_rmb(uint8_t weak_barriers) 46 { 47 if (weak_barriers) 48 rte_smp_rmb(); 49 else 50 rte_cio_rmb(); 51 } 52 53 static inline void 54 virtio_wmb(uint8_t weak_barriers) 55 { 56 if (weak_barriers) 57 rte_smp_wmb(); 58 else 59 rte_cio_wmb(); 60 } 61 62 static inline uint16_t 63 virtqueue_fetch_flags_packed(struct vring_packed_desc *dp, 64 uint8_t weak_barriers) 65 { 66 uint16_t flags; 67 68 if (weak_barriers) { 69 /* x86 prefers to using rte_smp_rmb over __atomic_load_n as it reports 70 * a better perf(~1.5%), which comes from the saved branch by the compiler. 71 * The if and else branch are identical with the smp and cio barriers both 72 * defined as compiler barriers on x86. 73 */ 74 #ifdef RTE_ARCH_X86_64 75 flags = dp->flags; 76 rte_smp_rmb(); 77 #else 78 flags = __atomic_load_n(&dp->flags, __ATOMIC_ACQUIRE); 79 #endif 80 } else { 81 flags = dp->flags; 82 rte_cio_rmb(); 83 } 84 85 return flags; 86 } 87 88 static inline void 89 virtqueue_store_flags_packed(struct vring_packed_desc *dp, 90 uint16_t flags, uint8_t weak_barriers) 91 { 92 if (weak_barriers) { 93 /* x86 prefers to using rte_smp_wmb over __atomic_store_n as it reports 94 * a better perf(~1.5%), which comes from the saved branch by the compiler. 95 * The if and else branch are identical with the smp and cio barriers both 96 * defined as compiler barriers on x86. 97 */ 98 #ifdef RTE_ARCH_X86_64 99 rte_smp_wmb(); 100 dp->flags = flags; 101 #else 102 __atomic_store_n(&dp->flags, flags, __ATOMIC_RELEASE); 103 #endif 104 } else { 105 rte_cio_wmb(); 106 dp->flags = flags; 107 } 108 } 109 #ifdef RTE_PMD_PACKET_PREFETCH 110 #define rte_packet_prefetch(p) rte_prefetch1(p) 111 #else 112 #define rte_packet_prefetch(p) do {} while(0) 113 #endif 114 115 #define VIRTQUEUE_MAX_NAME_SZ 32 116 117 #ifdef RTE_VIRTIO_USER 118 /** 119 * Return the physical address (or virtual address in case of 120 * virtio-user) of mbuf data buffer. 121 * 122 * The address is firstly casted to the word size (sizeof(uintptr_t)) 123 * before casting it to uint64_t. This is to make it work with different 124 * combination of word size (64 bit and 32 bit) and virtio device 125 * (virtio-pci and virtio-user). 126 */ 127 #define VIRTIO_MBUF_ADDR(mb, vq) \ 128 ((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->offset))) 129 #else 130 #define VIRTIO_MBUF_ADDR(mb, vq) ((mb)->buf_iova) 131 #endif 132 133 /** 134 * Return the physical address (or virtual address in case of 135 * virtio-user) of mbuf data buffer, taking care of mbuf data offset 136 */ 137 #define VIRTIO_MBUF_DATA_DMA_ADDR(mb, vq) \ 138 (VIRTIO_MBUF_ADDR(mb, vq) + (mb)->data_off) 139 140 #define VTNET_SQ_RQ_QUEUE_IDX 0 141 #define VTNET_SQ_TQ_QUEUE_IDX 1 142 #define VTNET_SQ_CQ_QUEUE_IDX 2 143 144 enum { VTNET_RQ = 0, VTNET_TQ = 1, VTNET_CQ = 2 }; 145 /** 146 * The maximum virtqueue size is 2^15. Use that value as the end of 147 * descriptor chain terminator since it will never be a valid index 148 * in the descriptor table. This is used to verify we are correctly 149 * handling vq_free_cnt. 150 */ 151 #define VQ_RING_DESC_CHAIN_END 32768 152 153 /** 154 * Control the RX mode, ie. promiscuous, allmulti, etc... 155 * All commands require an "out" sg entry containing a 1 byte 156 * state value, zero = disable, non-zero = enable. Commands 157 * 0 and 1 are supported with the VIRTIO_NET_F_CTRL_RX feature. 158 * Commands 2-5 are added with VIRTIO_NET_F_CTRL_RX_EXTRA. 159 */ 160 #define VIRTIO_NET_CTRL_RX 0 161 #define VIRTIO_NET_CTRL_RX_PROMISC 0 162 #define VIRTIO_NET_CTRL_RX_ALLMULTI 1 163 #define VIRTIO_NET_CTRL_RX_ALLUNI 2 164 #define VIRTIO_NET_CTRL_RX_NOMULTI 3 165 #define VIRTIO_NET_CTRL_RX_NOUNI 4 166 #define VIRTIO_NET_CTRL_RX_NOBCAST 5 167 168 /** 169 * Control the MAC 170 * 171 * The MAC filter table is managed by the hypervisor, the guest should 172 * assume the size is infinite. Filtering should be considered 173 * non-perfect, ie. based on hypervisor resources, the guest may 174 * received packets from sources not specified in the filter list. 175 * 176 * In addition to the class/cmd header, the TABLE_SET command requires 177 * two out scatterlists. Each contains a 4 byte count of entries followed 178 * by a concatenated byte stream of the ETH_ALEN MAC addresses. The 179 * first sg list contains unicast addresses, the second is for multicast. 180 * This functionality is present if the VIRTIO_NET_F_CTRL_RX feature 181 * is available. 182 * 183 * The ADDR_SET command requests one out scatterlist, it contains a 184 * 6 bytes MAC address. This functionality is present if the 185 * VIRTIO_NET_F_CTRL_MAC_ADDR feature is available. 186 */ 187 struct virtio_net_ctrl_mac { 188 uint32_t entries; 189 uint8_t macs[][RTE_ETHER_ADDR_LEN]; 190 } __rte_packed; 191 192 #define VIRTIO_NET_CTRL_MAC 1 193 #define VIRTIO_NET_CTRL_MAC_TABLE_SET 0 194 #define VIRTIO_NET_CTRL_MAC_ADDR_SET 1 195 196 /** 197 * Control VLAN filtering 198 * 199 * The VLAN filter table is controlled via a simple ADD/DEL interface. 200 * VLAN IDs not added may be filtered by the hypervisor. Del is the 201 * opposite of add. Both commands expect an out entry containing a 2 202 * byte VLAN ID. VLAN filtering is available with the 203 * VIRTIO_NET_F_CTRL_VLAN feature bit. 204 */ 205 #define VIRTIO_NET_CTRL_VLAN 2 206 #define VIRTIO_NET_CTRL_VLAN_ADD 0 207 #define VIRTIO_NET_CTRL_VLAN_DEL 1 208 209 /* 210 * Control link announce acknowledgement 211 * 212 * The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that 213 * driver has recevied the notification; device would clear the 214 * VIRTIO_NET_S_ANNOUNCE bit in the status field after it receives 215 * this command. 216 */ 217 #define VIRTIO_NET_CTRL_ANNOUNCE 3 218 #define VIRTIO_NET_CTRL_ANNOUNCE_ACK 0 219 220 struct virtio_net_ctrl_hdr { 221 uint8_t class; 222 uint8_t cmd; 223 } __rte_packed; 224 225 typedef uint8_t virtio_net_ctrl_ack; 226 227 #define VIRTIO_NET_OK 0 228 #define VIRTIO_NET_ERR 1 229 230 #define VIRTIO_MAX_CTRL_DATA 2048 231 232 struct virtio_pmd_ctrl { 233 struct virtio_net_ctrl_hdr hdr; 234 virtio_net_ctrl_ack status; 235 uint8_t data[VIRTIO_MAX_CTRL_DATA]; 236 }; 237 238 struct vq_desc_extra { 239 void *cookie; 240 uint16_t ndescs; 241 uint16_t next; 242 }; 243 244 struct virtqueue { 245 struct virtio_hw *hw; /**< virtio_hw structure pointer. */ 246 union { 247 struct { 248 /**< vring keeping desc, used and avail */ 249 struct vring ring; 250 } vq_split; 251 252 struct { 253 /**< vring keeping descs and events */ 254 struct vring_packed ring; 255 bool used_wrap_counter; 256 uint16_t cached_flags; /**< cached flags for descs */ 257 uint16_t event_flags_shadow; 258 } vq_packed; 259 }; 260 261 uint16_t vq_used_cons_idx; /**< last consumed descriptor */ 262 uint16_t vq_nentries; /**< vring desc numbers */ 263 uint16_t vq_free_cnt; /**< num of desc available */ 264 uint16_t vq_avail_idx; /**< sync until needed */ 265 uint16_t vq_free_thresh; /**< free threshold */ 266 267 void *vq_ring_virt_mem; /**< linear address of vring*/ 268 unsigned int vq_ring_size; 269 270 union { 271 struct virtnet_rx rxq; 272 struct virtnet_tx txq; 273 struct virtnet_ctl cq; 274 }; 275 276 rte_iova_t vq_ring_mem; /**< physical address of vring, 277 * or virtual address for virtio_user. */ 278 279 /** 280 * Head of the free chain in the descriptor table. If 281 * there are no free descriptors, this will be set to 282 * VQ_RING_DESC_CHAIN_END. 283 */ 284 uint16_t vq_desc_head_idx; 285 uint16_t vq_desc_tail_idx; 286 uint16_t vq_queue_index; /**< PCI queue index */ 287 uint16_t offset; /**< relative offset to obtain addr in mbuf */ 288 uint16_t *notify_addr; 289 struct rte_mbuf **sw_ring; /**< RX software ring. */ 290 struct vq_desc_extra vq_descx[0]; 291 }; 292 293 /* If multiqueue is provided by host, then we suppport it. */ 294 #define VIRTIO_NET_CTRL_MQ 4 295 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET 0 296 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN 1 297 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX 0x8000 298 299 /** 300 * This is the first element of the scatter-gather list. If you don't 301 * specify GSO or CSUM features, you can simply ignore the header. 302 */ 303 struct virtio_net_hdr { 304 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /**< Use csum_start,csum_offset*/ 305 #define VIRTIO_NET_HDR_F_DATA_VALID 2 /**< Checksum is valid */ 306 uint8_t flags; 307 #define VIRTIO_NET_HDR_GSO_NONE 0 /**< Not a GSO frame */ 308 #define VIRTIO_NET_HDR_GSO_TCPV4 1 /**< GSO frame, IPv4 TCP (TSO) */ 309 #define VIRTIO_NET_HDR_GSO_UDP 3 /**< GSO frame, IPv4 UDP (UFO) */ 310 #define VIRTIO_NET_HDR_GSO_TCPV6 4 /**< GSO frame, IPv6 TCP */ 311 #define VIRTIO_NET_HDR_GSO_ECN 0x80 /**< TCP has ECN set */ 312 uint8_t gso_type; 313 uint16_t hdr_len; /**< Ethernet + IP + tcp/udp hdrs */ 314 uint16_t gso_size; /**< Bytes to append to hdr_len per frame */ 315 uint16_t csum_start; /**< Position to start checksumming from */ 316 uint16_t csum_offset; /**< Offset after that to place checksum */ 317 }; 318 319 /** 320 * This is the version of the header to use when the MRG_RXBUF 321 * feature has been negotiated. 322 */ 323 struct virtio_net_hdr_mrg_rxbuf { 324 struct virtio_net_hdr hdr; 325 uint16_t num_buffers; /**< Number of merged rx buffers */ 326 }; 327 328 /* Region reserved to allow for transmit header and indirect ring */ 329 #define VIRTIO_MAX_TX_INDIRECT 8 330 struct virtio_tx_region { 331 struct virtio_net_hdr_mrg_rxbuf tx_hdr; 332 struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT] 333 __rte_aligned(16); 334 }; 335 336 static inline int 337 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq) 338 { 339 uint16_t used, avail, flags; 340 341 flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers); 342 used = !!(flags & VRING_PACKED_DESC_F_USED); 343 avail = !!(flags & VRING_PACKED_DESC_F_AVAIL); 344 345 return avail == used && used == vq->vq_packed.used_wrap_counter; 346 } 347 348 static inline void 349 vring_desc_init_packed(struct virtqueue *vq, int n) 350 { 351 int i; 352 for (i = 0; i < n - 1; i++) { 353 vq->vq_packed.ring.desc[i].id = i; 354 vq->vq_descx[i].next = i + 1; 355 } 356 vq->vq_packed.ring.desc[i].id = i; 357 vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END; 358 } 359 360 /* Chain all the descriptors in the ring with an END */ 361 static inline void 362 vring_desc_init_split(struct vring_desc *dp, uint16_t n) 363 { 364 uint16_t i; 365 366 for (i = 0; i < n - 1; i++) 367 dp[i].next = (uint16_t)(i + 1); 368 dp[i].next = VQ_RING_DESC_CHAIN_END; 369 } 370 371 /** 372 * Tell the backend not to interrupt us. Implementation for packed virtqueues. 373 */ 374 static inline void 375 virtqueue_disable_intr_packed(struct virtqueue *vq) 376 { 377 if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) { 378 vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE; 379 vq->vq_packed.ring.driver->desc_event_flags = 380 vq->vq_packed.event_flags_shadow; 381 } 382 } 383 384 /** 385 * Tell the backend not to interrupt us. Implementation for split virtqueues. 386 */ 387 static inline void 388 virtqueue_disable_intr_split(struct virtqueue *vq) 389 { 390 vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT; 391 } 392 393 /** 394 * Tell the backend not to interrupt us. 395 */ 396 static inline void 397 virtqueue_disable_intr(struct virtqueue *vq) 398 { 399 if (vtpci_packed_queue(vq->hw)) 400 virtqueue_disable_intr_packed(vq); 401 else 402 virtqueue_disable_intr_split(vq); 403 } 404 405 /** 406 * Tell the backend to interrupt. Implementation for packed virtqueues. 407 */ 408 static inline void 409 virtqueue_enable_intr_packed(struct virtqueue *vq) 410 { 411 if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) { 412 vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE; 413 vq->vq_packed.ring.driver->desc_event_flags = 414 vq->vq_packed.event_flags_shadow; 415 } 416 } 417 418 /** 419 * Tell the backend to interrupt. Implementation for split virtqueues. 420 */ 421 static inline void 422 virtqueue_enable_intr_split(struct virtqueue *vq) 423 { 424 vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT); 425 } 426 427 /** 428 * Tell the backend to interrupt us. 429 */ 430 static inline void 431 virtqueue_enable_intr(struct virtqueue *vq) 432 { 433 if (vtpci_packed_queue(vq->hw)) 434 virtqueue_enable_intr_packed(vq); 435 else 436 virtqueue_enable_intr_split(vq); 437 } 438 439 /** 440 * Dump virtqueue internal structures, for debug purpose only. 441 */ 442 void virtqueue_dump(struct virtqueue *vq); 443 /** 444 * Get all mbufs to be freed. 445 */ 446 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq); 447 448 /* Flush the elements in the used ring. */ 449 void virtqueue_rxvq_flush(struct virtqueue *vq); 450 451 int virtqueue_rxvq_reset_packed(struct virtqueue *vq); 452 453 int virtqueue_txvq_reset_packed(struct virtqueue *vq); 454 455 static inline int 456 virtqueue_full(const struct virtqueue *vq) 457 { 458 return vq->vq_free_cnt == 0; 459 } 460 461 static inline int 462 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vtpci_queue_idx) 463 { 464 if (vtpci_queue_idx == hw->max_queue_pairs * 2) 465 return VTNET_CQ; 466 else if (vtpci_queue_idx % 2 == 0) 467 return VTNET_RQ; 468 else 469 return VTNET_TQ; 470 } 471 472 /* virtqueue_nused has load-acquire or rte_cio_rmb insed */ 473 static inline uint16_t 474 virtqueue_nused(const struct virtqueue *vq) 475 { 476 uint16_t idx; 477 478 if (vq->hw->weak_barriers) { 479 /** 480 * x86 prefers to using rte_smp_rmb over __atomic_load_n as it 481 * reports a slightly better perf, which comes from the saved 482 * branch by the compiler. 483 * The if and else branches are identical with the smp and cio 484 * barriers both defined as compiler barriers on x86. 485 */ 486 #ifdef RTE_ARCH_X86_64 487 idx = vq->vq_split.ring.used->idx; 488 rte_smp_rmb(); 489 #else 490 idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, 491 __ATOMIC_ACQUIRE); 492 #endif 493 } else { 494 idx = vq->vq_split.ring.used->idx; 495 rte_cio_rmb(); 496 } 497 return idx - vq->vq_used_cons_idx; 498 } 499 500 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx); 501 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx); 502 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, 503 uint16_t num); 504 505 static inline void 506 vq_update_avail_idx(struct virtqueue *vq) 507 { 508 virtio_wmb(vq->hw->weak_barriers); 509 vq->vq_split.ring.avail->idx = vq->vq_avail_idx; 510 } 511 512 static inline void 513 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx) 514 { 515 uint16_t avail_idx; 516 /* 517 * Place the head of the descriptor chain into the next slot and make 518 * it usable to the host. The chain is made available now rather than 519 * deferring to virtqueue_notify() in the hopes that if the host is 520 * currently running on another CPU, we can keep it processing the new 521 * descriptor. 522 */ 523 avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1)); 524 if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx)) 525 vq->vq_split.ring.avail->ring[avail_idx] = desc_idx; 526 vq->vq_avail_idx++; 527 } 528 529 static inline int 530 virtqueue_kick_prepare(struct virtqueue *vq) 531 { 532 /* 533 * Ensure updated avail->idx is visible to vhost before reading 534 * the used->flags. 535 */ 536 virtio_mb(vq->hw->weak_barriers); 537 return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY); 538 } 539 540 static inline int 541 virtqueue_kick_prepare_packed(struct virtqueue *vq) 542 { 543 uint16_t flags; 544 545 /* 546 * Ensure updated data is visible to vhost before reading the flags. 547 */ 548 virtio_mb(vq->hw->weak_barriers); 549 flags = vq->vq_packed.ring.device->desc_event_flags; 550 551 return flags != RING_EVENT_FLAGS_DISABLE; 552 } 553 554 /* 555 * virtqueue_kick_prepare*() or the virtio_wmb() should be called 556 * before this function to be sure that all the data is visible to vhost. 557 */ 558 static inline void 559 virtqueue_notify(struct virtqueue *vq) 560 { 561 VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq); 562 } 563 564 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP 565 #define VIRTQUEUE_DUMP(vq) do { \ 566 uint16_t used_idx, nused; \ 567 used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \ 568 __ATOMIC_RELAXED); \ 569 nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \ 570 if (vtpci_packed_queue((vq)->hw)) { \ 571 PMD_INIT_LOG(DEBUG, \ 572 "VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \ 573 " cached_flags=0x%x; used_wrap_counter=%d", \ 574 (vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \ 575 (vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \ 576 (vq)->vq_packed.used_wrap_counter); \ 577 break; \ 578 } \ 579 PMD_INIT_LOG(DEBUG, \ 580 "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \ 581 " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \ 582 " avail.flags=0x%x; used.flags=0x%x", \ 583 (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \ 584 (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \ 585 __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \ 586 (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \ 587 } while (0) 588 #else 589 #define VIRTQUEUE_DUMP(vq) do { } while (0) 590 #endif 591 592 /* avoid write operation when necessary, to lessen cache issues */ 593 #define ASSIGN_UNLESS_EQUAL(var, val) do { \ 594 typeof(var) var_ = (var); \ 595 typeof(val) val_ = (val); \ 596 if ((var_) != (val_)) \ 597 (var_) = (val_); \ 598 } while (0) 599 600 #define virtqueue_clear_net_hdr(hdr) do { \ 601 typeof(hdr) hdr_ = (hdr); \ 602 ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0); \ 603 ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0); \ 604 ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0); \ 605 ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0); \ 606 ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0); \ 607 ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0); \ 608 } while (0) 609 610 static inline void 611 virtqueue_xmit_offload(struct virtio_net_hdr *hdr, 612 struct rte_mbuf *cookie, 613 bool offload) 614 { 615 if (offload) { 616 if (cookie->ol_flags & PKT_TX_TCP_SEG) 617 cookie->ol_flags |= PKT_TX_TCP_CKSUM; 618 619 switch (cookie->ol_flags & PKT_TX_L4_MASK) { 620 case PKT_TX_UDP_CKSUM: 621 hdr->csum_start = cookie->l2_len + cookie->l3_len; 622 hdr->csum_offset = offsetof(struct rte_udp_hdr, 623 dgram_cksum); 624 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; 625 break; 626 627 case PKT_TX_TCP_CKSUM: 628 hdr->csum_start = cookie->l2_len + cookie->l3_len; 629 hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum); 630 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; 631 break; 632 633 default: 634 ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0); 635 ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0); 636 ASSIGN_UNLESS_EQUAL(hdr->flags, 0); 637 break; 638 } 639 640 /* TCP Segmentation Offload */ 641 if (cookie->ol_flags & PKT_TX_TCP_SEG) { 642 hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ? 643 VIRTIO_NET_HDR_GSO_TCPV6 : 644 VIRTIO_NET_HDR_GSO_TCPV4; 645 hdr->gso_size = cookie->tso_segsz; 646 hdr->hdr_len = 647 cookie->l2_len + 648 cookie->l3_len + 649 cookie->l4_len; 650 } else { 651 ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0); 652 ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0); 653 ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0); 654 } 655 } 656 } 657 658 static inline void 659 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie, 660 uint16_t needed, int can_push, int in_order) 661 { 662 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr; 663 struct vq_desc_extra *dxp; 664 struct virtqueue *vq = txvq->vq; 665 struct vring_packed_desc *start_dp, *head_dp; 666 uint16_t idx, id, head_idx, head_flags; 667 int16_t head_size = vq->hw->vtnet_hdr_size; 668 struct virtio_net_hdr *hdr; 669 uint16_t prev; 670 bool prepend_header = false; 671 672 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx; 673 674 dxp = &vq->vq_descx[id]; 675 dxp->ndescs = needed; 676 dxp->cookie = cookie; 677 678 head_idx = vq->vq_avail_idx; 679 idx = head_idx; 680 prev = head_idx; 681 start_dp = vq->vq_packed.ring.desc; 682 683 head_dp = &vq->vq_packed.ring.desc[idx]; 684 head_flags = cookie->next ? VRING_DESC_F_NEXT : 0; 685 head_flags |= vq->vq_packed.cached_flags; 686 687 if (can_push) { 688 /* prepend cannot fail, checked by caller */ 689 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *, 690 -head_size); 691 prepend_header = true; 692 693 /* if offload disabled, it is not zeroed below, do it now */ 694 if (!vq->hw->has_tx_offload) 695 virtqueue_clear_net_hdr(hdr); 696 } else { 697 /* setup first tx ring slot to point to header 698 * stored in reserved region. 699 */ 700 start_dp[idx].addr = txvq->virtio_net_hdr_mem + 701 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr); 702 start_dp[idx].len = vq->hw->vtnet_hdr_size; 703 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr; 704 idx++; 705 if (idx >= vq->vq_nentries) { 706 idx -= vq->vq_nentries; 707 vq->vq_packed.cached_flags ^= 708 VRING_PACKED_DESC_F_AVAIL_USED; 709 } 710 } 711 712 virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload); 713 714 do { 715 uint16_t flags; 716 717 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq); 718 start_dp[idx].len = cookie->data_len; 719 if (prepend_header) { 720 start_dp[idx].addr -= head_size; 721 start_dp[idx].len += head_size; 722 prepend_header = false; 723 } 724 725 if (likely(idx != head_idx)) { 726 flags = cookie->next ? VRING_DESC_F_NEXT : 0; 727 flags |= vq->vq_packed.cached_flags; 728 start_dp[idx].flags = flags; 729 } 730 prev = idx; 731 idx++; 732 if (idx >= vq->vq_nentries) { 733 idx -= vq->vq_nentries; 734 vq->vq_packed.cached_flags ^= 735 VRING_PACKED_DESC_F_AVAIL_USED; 736 } 737 } while ((cookie = cookie->next) != NULL); 738 739 start_dp[prev].id = id; 740 741 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed); 742 vq->vq_avail_idx = idx; 743 744 if (!in_order) { 745 vq->vq_desc_head_idx = dxp->next; 746 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) 747 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END; 748 } 749 750 virtqueue_store_flags_packed(head_dp, head_flags, 751 vq->hw->weak_barriers); 752 } 753 754 static void 755 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id) 756 { 757 struct vq_desc_extra *dxp; 758 759 dxp = &vq->vq_descx[id]; 760 vq->vq_free_cnt += dxp->ndescs; 761 762 if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) 763 vq->vq_desc_head_idx = id; 764 else 765 vq->vq_descx[vq->vq_desc_tail_idx].next = id; 766 767 vq->vq_desc_tail_idx = id; 768 dxp->next = VQ_RING_DESC_CHAIN_END; 769 } 770 771 static void 772 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num) 773 { 774 uint16_t used_idx, id, curr_id, free_cnt = 0; 775 uint16_t size = vq->vq_nentries; 776 struct vring_packed_desc *desc = vq->vq_packed.ring.desc; 777 struct vq_desc_extra *dxp; 778 779 used_idx = vq->vq_used_cons_idx; 780 /* desc_is_used has a load-acquire or rte_cio_rmb inside 781 * and wait for used desc in virtqueue. 782 */ 783 while (num > 0 && desc_is_used(&desc[used_idx], vq)) { 784 id = desc[used_idx].id; 785 do { 786 curr_id = used_idx; 787 dxp = &vq->vq_descx[used_idx]; 788 used_idx += dxp->ndescs; 789 free_cnt += dxp->ndescs; 790 num -= dxp->ndescs; 791 if (used_idx >= size) { 792 used_idx -= size; 793 vq->vq_packed.used_wrap_counter ^= 1; 794 } 795 if (dxp->cookie != NULL) { 796 rte_pktmbuf_free(dxp->cookie); 797 dxp->cookie = NULL; 798 } 799 } while (curr_id != id); 800 } 801 vq->vq_used_cons_idx = used_idx; 802 vq->vq_free_cnt += free_cnt; 803 } 804 805 static void 806 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num) 807 { 808 uint16_t used_idx, id; 809 uint16_t size = vq->vq_nentries; 810 struct vring_packed_desc *desc = vq->vq_packed.ring.desc; 811 struct vq_desc_extra *dxp; 812 813 used_idx = vq->vq_used_cons_idx; 814 /* desc_is_used has a load-acquire or rte_cio_rmb inside 815 * and wait for used desc in virtqueue. 816 */ 817 while (num-- && desc_is_used(&desc[used_idx], vq)) { 818 id = desc[used_idx].id; 819 dxp = &vq->vq_descx[id]; 820 vq->vq_used_cons_idx += dxp->ndescs; 821 if (vq->vq_used_cons_idx >= size) { 822 vq->vq_used_cons_idx -= size; 823 vq->vq_packed.used_wrap_counter ^= 1; 824 } 825 vq_ring_free_id_packed(vq, id); 826 if (dxp->cookie != NULL) { 827 rte_pktmbuf_free(dxp->cookie); 828 dxp->cookie = NULL; 829 } 830 used_idx = vq->vq_used_cons_idx; 831 } 832 } 833 834 /* Cleanup from completed transmits. */ 835 static inline void 836 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order) 837 { 838 if (in_order) 839 virtio_xmit_cleanup_inorder_packed(vq, num); 840 else 841 virtio_xmit_cleanup_normal_packed(vq, num); 842 } 843 844 static inline void 845 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num) 846 { 847 uint16_t i, used_idx, desc_idx; 848 for (i = 0; i < num; i++) { 849 struct vring_used_elem *uep; 850 struct vq_desc_extra *dxp; 851 852 used_idx = (uint16_t)(vq->vq_used_cons_idx & 853 (vq->vq_nentries - 1)); 854 uep = &vq->vq_split.ring.used->ring[used_idx]; 855 856 desc_idx = (uint16_t)uep->id; 857 dxp = &vq->vq_descx[desc_idx]; 858 vq->vq_used_cons_idx++; 859 vq_ring_free_chain(vq, desc_idx); 860 861 if (dxp->cookie != NULL) { 862 rte_pktmbuf_free(dxp->cookie); 863 dxp->cookie = NULL; 864 } 865 } 866 } 867 868 /* Cleanup from completed inorder transmits. */ 869 static __rte_always_inline void 870 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num) 871 { 872 uint16_t i, idx = vq->vq_used_cons_idx; 873 int16_t free_cnt = 0; 874 struct vq_desc_extra *dxp = NULL; 875 876 if (unlikely(num == 0)) 877 return; 878 879 for (i = 0; i < num; i++) { 880 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)]; 881 free_cnt += dxp->ndescs; 882 if (dxp->cookie != NULL) { 883 rte_pktmbuf_free(dxp->cookie); 884 dxp->cookie = NULL; 885 } 886 } 887 888 vq->vq_free_cnt += free_cnt; 889 vq->vq_used_cons_idx = idx; 890 } 891 #endif /* _VIRTQUEUE_H_ */ 892