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 if (vq->hw->weak_barriers) { 509 /* x86 prefers to using rte_smp_wmb over __atomic_store_n as 510 * it reports a slightly better perf, which comes from the 511 * saved branch by the compiler. 512 * The if and else branches are identical with the smp and 513 * cio barriers both defined as compiler barriers on x86. 514 */ 515 #ifdef RTE_ARCH_X86_64 516 rte_smp_wmb(); 517 vq->vq_split.ring.avail->idx = vq->vq_avail_idx; 518 #else 519 __atomic_store_n(&vq->vq_split.ring.avail->idx, 520 vq->vq_avail_idx, __ATOMIC_RELEASE); 521 #endif 522 } else { 523 rte_cio_wmb(); 524 vq->vq_split.ring.avail->idx = vq->vq_avail_idx; 525 } 526 } 527 528 static inline void 529 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx) 530 { 531 uint16_t avail_idx; 532 /* 533 * Place the head of the descriptor chain into the next slot and make 534 * it usable to the host. The chain is made available now rather than 535 * deferring to virtqueue_notify() in the hopes that if the host is 536 * currently running on another CPU, we can keep it processing the new 537 * descriptor. 538 */ 539 avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1)); 540 if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx)) 541 vq->vq_split.ring.avail->ring[avail_idx] = desc_idx; 542 vq->vq_avail_idx++; 543 } 544 545 static inline int 546 virtqueue_kick_prepare(struct virtqueue *vq) 547 { 548 /* 549 * Ensure updated avail->idx is visible to vhost before reading 550 * the used->flags. 551 */ 552 virtio_mb(vq->hw->weak_barriers); 553 return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY); 554 } 555 556 static inline int 557 virtqueue_kick_prepare_packed(struct virtqueue *vq) 558 { 559 uint16_t flags; 560 561 /* 562 * Ensure updated data is visible to vhost before reading the flags. 563 */ 564 virtio_mb(vq->hw->weak_barriers); 565 flags = vq->vq_packed.ring.device->desc_event_flags; 566 567 return flags != RING_EVENT_FLAGS_DISABLE; 568 } 569 570 /* 571 * virtqueue_kick_prepare*() or the virtio_wmb() should be called 572 * before this function to be sure that all the data is visible to vhost. 573 */ 574 static inline void 575 virtqueue_notify(struct virtqueue *vq) 576 { 577 VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq); 578 } 579 580 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP 581 #define VIRTQUEUE_DUMP(vq) do { \ 582 uint16_t used_idx, nused; \ 583 used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \ 584 __ATOMIC_RELAXED); \ 585 nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \ 586 if (vtpci_packed_queue((vq)->hw)) { \ 587 PMD_INIT_LOG(DEBUG, \ 588 "VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \ 589 " cached_flags=0x%x; used_wrap_counter=%d", \ 590 (vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \ 591 (vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \ 592 (vq)->vq_packed.used_wrap_counter); \ 593 break; \ 594 } \ 595 PMD_INIT_LOG(DEBUG, \ 596 "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \ 597 " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \ 598 " avail.flags=0x%x; used.flags=0x%x", \ 599 (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \ 600 (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \ 601 __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \ 602 (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \ 603 } while (0) 604 #else 605 #define VIRTQUEUE_DUMP(vq) do { } while (0) 606 #endif 607 608 /* avoid write operation when necessary, to lessen cache issues */ 609 #define ASSIGN_UNLESS_EQUAL(var, val) do { \ 610 typeof(var) var_ = (var); \ 611 typeof(val) val_ = (val); \ 612 if ((var_) != (val_)) \ 613 (var_) = (val_); \ 614 } while (0) 615 616 #define virtqueue_clear_net_hdr(hdr) do { \ 617 typeof(hdr) hdr_ = (hdr); \ 618 ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0); \ 619 ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0); \ 620 ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0); \ 621 ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0); \ 622 ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0); \ 623 ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0); \ 624 } while (0) 625 626 static inline void 627 virtqueue_xmit_offload(struct virtio_net_hdr *hdr, 628 struct rte_mbuf *cookie, 629 bool offload) 630 { 631 if (offload) { 632 if (cookie->ol_flags & PKT_TX_TCP_SEG) 633 cookie->ol_flags |= PKT_TX_TCP_CKSUM; 634 635 switch (cookie->ol_flags & PKT_TX_L4_MASK) { 636 case PKT_TX_UDP_CKSUM: 637 hdr->csum_start = cookie->l2_len + cookie->l3_len; 638 hdr->csum_offset = offsetof(struct rte_udp_hdr, 639 dgram_cksum); 640 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; 641 break; 642 643 case PKT_TX_TCP_CKSUM: 644 hdr->csum_start = cookie->l2_len + cookie->l3_len; 645 hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum); 646 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM; 647 break; 648 649 default: 650 ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0); 651 ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0); 652 ASSIGN_UNLESS_EQUAL(hdr->flags, 0); 653 break; 654 } 655 656 /* TCP Segmentation Offload */ 657 if (cookie->ol_flags & PKT_TX_TCP_SEG) { 658 hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ? 659 VIRTIO_NET_HDR_GSO_TCPV6 : 660 VIRTIO_NET_HDR_GSO_TCPV4; 661 hdr->gso_size = cookie->tso_segsz; 662 hdr->hdr_len = 663 cookie->l2_len + 664 cookie->l3_len + 665 cookie->l4_len; 666 } else { 667 ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0); 668 ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0); 669 ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0); 670 } 671 } 672 } 673 674 static inline void 675 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie, 676 uint16_t needed, int can_push, int in_order) 677 { 678 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr; 679 struct vq_desc_extra *dxp; 680 struct virtqueue *vq = txvq->vq; 681 struct vring_packed_desc *start_dp, *head_dp; 682 uint16_t idx, id, head_idx, head_flags; 683 int16_t head_size = vq->hw->vtnet_hdr_size; 684 struct virtio_net_hdr *hdr; 685 uint16_t prev; 686 bool prepend_header = false; 687 688 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx; 689 690 dxp = &vq->vq_descx[id]; 691 dxp->ndescs = needed; 692 dxp->cookie = cookie; 693 694 head_idx = vq->vq_avail_idx; 695 idx = head_idx; 696 prev = head_idx; 697 start_dp = vq->vq_packed.ring.desc; 698 699 head_dp = &vq->vq_packed.ring.desc[idx]; 700 head_flags = cookie->next ? VRING_DESC_F_NEXT : 0; 701 head_flags |= vq->vq_packed.cached_flags; 702 703 if (can_push) { 704 /* prepend cannot fail, checked by caller */ 705 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *, 706 -head_size); 707 prepend_header = true; 708 709 /* if offload disabled, it is not zeroed below, do it now */ 710 if (!vq->hw->has_tx_offload) 711 virtqueue_clear_net_hdr(hdr); 712 } else { 713 /* setup first tx ring slot to point to header 714 * stored in reserved region. 715 */ 716 start_dp[idx].addr = txvq->virtio_net_hdr_mem + 717 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr); 718 start_dp[idx].len = vq->hw->vtnet_hdr_size; 719 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr; 720 idx++; 721 if (idx >= vq->vq_nentries) { 722 idx -= vq->vq_nentries; 723 vq->vq_packed.cached_flags ^= 724 VRING_PACKED_DESC_F_AVAIL_USED; 725 } 726 } 727 728 virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload); 729 730 do { 731 uint16_t flags; 732 733 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq); 734 start_dp[idx].len = cookie->data_len; 735 if (prepend_header) { 736 start_dp[idx].addr -= head_size; 737 start_dp[idx].len += head_size; 738 prepend_header = false; 739 } 740 741 if (likely(idx != head_idx)) { 742 flags = cookie->next ? VRING_DESC_F_NEXT : 0; 743 flags |= vq->vq_packed.cached_flags; 744 start_dp[idx].flags = flags; 745 } 746 prev = idx; 747 idx++; 748 if (idx >= vq->vq_nentries) { 749 idx -= vq->vq_nentries; 750 vq->vq_packed.cached_flags ^= 751 VRING_PACKED_DESC_F_AVAIL_USED; 752 } 753 } while ((cookie = cookie->next) != NULL); 754 755 start_dp[prev].id = id; 756 757 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed); 758 vq->vq_avail_idx = idx; 759 760 if (!in_order) { 761 vq->vq_desc_head_idx = dxp->next; 762 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END) 763 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END; 764 } 765 766 virtqueue_store_flags_packed(head_dp, head_flags, 767 vq->hw->weak_barriers); 768 } 769 770 static void 771 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id) 772 { 773 struct vq_desc_extra *dxp; 774 775 dxp = &vq->vq_descx[id]; 776 vq->vq_free_cnt += dxp->ndescs; 777 778 if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) 779 vq->vq_desc_head_idx = id; 780 else 781 vq->vq_descx[vq->vq_desc_tail_idx].next = id; 782 783 vq->vq_desc_tail_idx = id; 784 dxp->next = VQ_RING_DESC_CHAIN_END; 785 } 786 787 static void 788 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num) 789 { 790 uint16_t used_idx, id, curr_id, free_cnt = 0; 791 uint16_t size = vq->vq_nentries; 792 struct vring_packed_desc *desc = vq->vq_packed.ring.desc; 793 struct vq_desc_extra *dxp; 794 795 used_idx = vq->vq_used_cons_idx; 796 /* desc_is_used has a load-acquire or rte_cio_rmb inside 797 * and wait for used desc in virtqueue. 798 */ 799 while (num > 0 && desc_is_used(&desc[used_idx], vq)) { 800 id = desc[used_idx].id; 801 do { 802 curr_id = used_idx; 803 dxp = &vq->vq_descx[used_idx]; 804 used_idx += dxp->ndescs; 805 free_cnt += dxp->ndescs; 806 num -= dxp->ndescs; 807 if (used_idx >= size) { 808 used_idx -= size; 809 vq->vq_packed.used_wrap_counter ^= 1; 810 } 811 if (dxp->cookie != NULL) { 812 rte_pktmbuf_free(dxp->cookie); 813 dxp->cookie = NULL; 814 } 815 } while (curr_id != id); 816 } 817 vq->vq_used_cons_idx = used_idx; 818 vq->vq_free_cnt += free_cnt; 819 } 820 821 static void 822 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num) 823 { 824 uint16_t used_idx, id; 825 uint16_t size = vq->vq_nentries; 826 struct vring_packed_desc *desc = vq->vq_packed.ring.desc; 827 struct vq_desc_extra *dxp; 828 829 used_idx = vq->vq_used_cons_idx; 830 /* desc_is_used has a load-acquire or rte_cio_rmb inside 831 * and wait for used desc in virtqueue. 832 */ 833 while (num-- && desc_is_used(&desc[used_idx], vq)) { 834 id = desc[used_idx].id; 835 dxp = &vq->vq_descx[id]; 836 vq->vq_used_cons_idx += dxp->ndescs; 837 if (vq->vq_used_cons_idx >= size) { 838 vq->vq_used_cons_idx -= size; 839 vq->vq_packed.used_wrap_counter ^= 1; 840 } 841 vq_ring_free_id_packed(vq, id); 842 if (dxp->cookie != NULL) { 843 rte_pktmbuf_free(dxp->cookie); 844 dxp->cookie = NULL; 845 } 846 used_idx = vq->vq_used_cons_idx; 847 } 848 } 849 850 /* Cleanup from completed transmits. */ 851 static inline void 852 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order) 853 { 854 if (in_order) 855 virtio_xmit_cleanup_inorder_packed(vq, num); 856 else 857 virtio_xmit_cleanup_normal_packed(vq, num); 858 } 859 860 static inline void 861 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num) 862 { 863 uint16_t i, used_idx, desc_idx; 864 for (i = 0; i < num; i++) { 865 struct vring_used_elem *uep; 866 struct vq_desc_extra *dxp; 867 868 used_idx = (uint16_t)(vq->vq_used_cons_idx & 869 (vq->vq_nentries - 1)); 870 uep = &vq->vq_split.ring.used->ring[used_idx]; 871 872 desc_idx = (uint16_t)uep->id; 873 dxp = &vq->vq_descx[desc_idx]; 874 vq->vq_used_cons_idx++; 875 vq_ring_free_chain(vq, desc_idx); 876 877 if (dxp->cookie != NULL) { 878 rte_pktmbuf_free(dxp->cookie); 879 dxp->cookie = NULL; 880 } 881 } 882 } 883 884 /* Cleanup from completed inorder transmits. */ 885 static __rte_always_inline void 886 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num) 887 { 888 uint16_t i, idx = vq->vq_used_cons_idx; 889 int16_t free_cnt = 0; 890 struct vq_desc_extra *dxp = NULL; 891 892 if (unlikely(num == 0)) 893 return; 894 895 for (i = 0; i < num; i++) { 896 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)]; 897 free_cnt += dxp->ndescs; 898 if (dxp->cookie != NULL) { 899 rte_pktmbuf_free(dxp->cookie); 900 dxp->cookie = NULL; 901 } 902 } 903 904 vq->vq_free_cnt += free_cnt; 905 vq->vq_used_cons_idx = idx; 906 } 907 #endif /* _VIRTQUEUE_H_ */ 908