1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2016-2018 Microsoft Corporation 3 * Copyright(c) 2013-2016 Brocade Communications Systems, Inc. 4 * All rights reserved. 5 */ 6 7 #include <stdint.h> 8 #include <string.h> 9 #include <stdio.h> 10 #include <errno.h> 11 #include <unistd.h> 12 #include <strings.h> 13 #include <malloc.h> 14 15 #include <rte_ethdev.h> 16 #include <rte_memcpy.h> 17 #include <rte_string_fns.h> 18 #include <rte_memzone.h> 19 #include <rte_malloc.h> 20 #include <rte_atomic.h> 21 #include <rte_bitmap.h> 22 #include <rte_branch_prediction.h> 23 #include <rte_ether.h> 24 #include <rte_common.h> 25 #include <rte_errno.h> 26 #include <rte_memory.h> 27 #include <rte_eal.h> 28 #include <rte_dev.h> 29 #include <rte_net.h> 30 #include <rte_bus_vmbus.h> 31 #include <rte_spinlock.h> 32 33 #include "hn_logs.h" 34 #include "hn_var.h" 35 #include "hn_rndis.h" 36 #include "hn_nvs.h" 37 #include "ndis.h" 38 39 #define HN_NVS_SEND_MSG_SIZE \ 40 (sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis)) 41 42 #define HN_TXD_CACHE_SIZE 32 /* per cpu tx_descriptor pool cache */ 43 #define HN_TXCOPY_THRESHOLD 512 44 45 #define HN_RXCOPY_THRESHOLD 256 46 #define HN_RXQ_EVENT_DEFAULT 2048 47 48 struct hn_rxinfo { 49 uint32_t vlan_info; 50 uint32_t csum_info; 51 uint32_t hash_info; 52 uint32_t hash_value; 53 }; 54 55 #define HN_RXINFO_VLAN 0x0001 56 #define HN_RXINFO_CSUM 0x0002 57 #define HN_RXINFO_HASHINF 0x0004 58 #define HN_RXINFO_HASHVAL 0x0008 59 #define HN_RXINFO_ALL \ 60 (HN_RXINFO_VLAN | \ 61 HN_RXINFO_CSUM | \ 62 HN_RXINFO_HASHINF | \ 63 HN_RXINFO_HASHVAL) 64 65 #define HN_NDIS_VLAN_INFO_INVALID 0xffffffff 66 #define HN_NDIS_RXCSUM_INFO_INVALID 0 67 #define HN_NDIS_HASH_INFO_INVALID 0 68 69 /* 70 * Per-transmit book keeping. 71 * A slot in transmit ring (chim_index) is reserved for each transmit. 72 * 73 * There are two types of transmit: 74 * - buffered transmit where chimney buffer is used and RNDIS header 75 * is in the buffer. mbuf == NULL for this case. 76 * 77 * - direct transmit where RNDIS header is in the in rndis_pkt 78 * mbuf is freed after transmit. 79 * 80 * Descriptors come from per-port pool which is used 81 * to limit number of outstanding requests per device. 82 */ 83 struct hn_txdesc { 84 struct rte_mbuf *m; 85 86 uint16_t queue_id; 87 uint32_t chim_index; 88 uint32_t chim_size; 89 uint32_t data_size; 90 uint32_t packets; 91 92 struct rndis_packet_msg *rndis_pkt; 93 }; 94 95 #define HN_RNDIS_PKT_LEN \ 96 (sizeof(struct rndis_packet_msg) + \ 97 RNDIS_PKTINFO_SIZE(NDIS_HASH_VALUE_SIZE) + \ 98 RNDIS_PKTINFO_SIZE(NDIS_VLAN_INFO_SIZE) + \ 99 RNDIS_PKTINFO_SIZE(NDIS_LSO2_INFO_SIZE) + \ 100 RNDIS_PKTINFO_SIZE(NDIS_TXCSUM_INFO_SIZE)) 101 102 #define HN_RNDIS_PKT_ALIGNED RTE_ALIGN(HN_RNDIS_PKT_LEN, RTE_CACHE_LINE_SIZE) 103 104 /* Minimum space required for a packet */ 105 #define HN_PKTSIZE_MIN(align) \ 106 RTE_ALIGN(RTE_ETHER_MIN_LEN + HN_RNDIS_PKT_LEN, align) 107 108 #define DEFAULT_TX_FREE_THRESH 32 109 110 static void 111 hn_update_packet_stats(struct hn_stats *stats, const struct rte_mbuf *m) 112 { 113 uint32_t s = m->pkt_len; 114 const struct rte_ether_addr *ea; 115 116 if (s == 64) { 117 stats->size_bins[1]++; 118 } else if (s > 64 && s < 1024) { 119 uint32_t bin; 120 121 /* count zeros, and offset into correct bin */ 122 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5; 123 stats->size_bins[bin]++; 124 } else { 125 if (s < 64) 126 stats->size_bins[0]++; 127 else if (s < 1519) 128 stats->size_bins[6]++; 129 else 130 stats->size_bins[7]++; 131 } 132 133 ea = rte_pktmbuf_mtod(m, const struct rte_ether_addr *); 134 if (rte_is_multicast_ether_addr(ea)) { 135 if (rte_is_broadcast_ether_addr(ea)) 136 stats->broadcast++; 137 else 138 stats->multicast++; 139 } 140 } 141 142 static inline unsigned int hn_rndis_pktlen(const struct rndis_packet_msg *pkt) 143 { 144 return pkt->pktinfooffset + pkt->pktinfolen; 145 } 146 147 static inline uint32_t 148 hn_rndis_pktmsg_offset(uint32_t ofs) 149 { 150 return ofs - offsetof(struct rndis_packet_msg, dataoffset); 151 } 152 153 static void hn_txd_init(struct rte_mempool *mp __rte_unused, 154 void *opaque, void *obj, unsigned int idx) 155 { 156 struct hn_tx_queue *txq = opaque; 157 struct hn_txdesc *txd = obj; 158 159 memset(txd, 0, sizeof(*txd)); 160 161 txd->queue_id = txq->queue_id; 162 txd->chim_index = NVS_CHIM_IDX_INVALID; 163 txd->rndis_pkt = (struct rndis_packet_msg *)((char *)txq->tx_rndis 164 + idx * HN_RNDIS_PKT_ALIGNED); 165 } 166 167 int 168 hn_chim_init(struct rte_eth_dev *dev) 169 { 170 struct hn_data *hv = dev->data->dev_private; 171 uint32_t i, chim_bmp_size; 172 173 rte_spinlock_init(&hv->chim_lock); 174 chim_bmp_size = rte_bitmap_get_memory_footprint(hv->chim_cnt); 175 hv->chim_bmem = rte_zmalloc("hn_chim_bitmap", chim_bmp_size, 176 RTE_CACHE_LINE_SIZE); 177 if (hv->chim_bmem == NULL) { 178 PMD_INIT_LOG(ERR, "failed to allocate bitmap size %u", 179 chim_bmp_size); 180 return -1; 181 } 182 183 hv->chim_bmap = rte_bitmap_init(hv->chim_cnt, 184 hv->chim_bmem, chim_bmp_size); 185 if (hv->chim_bmap == NULL) { 186 PMD_INIT_LOG(ERR, "failed to init chim bitmap"); 187 return -1; 188 } 189 190 for (i = 0; i < hv->chim_cnt; i++) 191 rte_bitmap_set(hv->chim_bmap, i); 192 193 return 0; 194 } 195 196 void 197 hn_chim_uninit(struct rte_eth_dev *dev) 198 { 199 struct hn_data *hv = dev->data->dev_private; 200 201 rte_bitmap_free(hv->chim_bmap); 202 rte_free(hv->chim_bmem); 203 hv->chim_bmem = NULL; 204 } 205 206 static uint32_t hn_chim_alloc(struct hn_data *hv) 207 { 208 uint32_t index = NVS_CHIM_IDX_INVALID; 209 uint64_t slab = 0; 210 211 rte_spinlock_lock(&hv->chim_lock); 212 if (rte_bitmap_scan(hv->chim_bmap, &index, &slab)) { 213 index += rte_bsf64(slab); 214 rte_bitmap_clear(hv->chim_bmap, index); 215 } 216 rte_spinlock_unlock(&hv->chim_lock); 217 218 return index; 219 } 220 221 static void hn_chim_free(struct hn_data *hv, uint32_t chim_idx) 222 { 223 if (chim_idx >= hv->chim_cnt) { 224 PMD_DRV_LOG(ERR, "Invalid chimney index %u", chim_idx); 225 } else { 226 rte_spinlock_lock(&hv->chim_lock); 227 rte_bitmap_set(hv->chim_bmap, chim_idx); 228 rte_spinlock_unlock(&hv->chim_lock); 229 } 230 } 231 232 static void hn_reset_txagg(struct hn_tx_queue *txq) 233 { 234 txq->agg_szleft = txq->agg_szmax; 235 txq->agg_pktleft = txq->agg_pktmax; 236 txq->agg_txd = NULL; 237 txq->agg_prevpkt = NULL; 238 } 239 240 int 241 hn_dev_tx_queue_setup(struct rte_eth_dev *dev, 242 uint16_t queue_idx, uint16_t nb_desc, 243 unsigned int socket_id, 244 const struct rte_eth_txconf *tx_conf) 245 246 { 247 struct hn_data *hv = dev->data->dev_private; 248 struct hn_tx_queue *txq; 249 char name[RTE_MEMPOOL_NAMESIZE]; 250 uint32_t tx_free_thresh; 251 int err = -ENOMEM; 252 253 PMD_INIT_FUNC_TRACE(); 254 255 txq = rte_zmalloc_socket("HN_TXQ", sizeof(*txq), RTE_CACHE_LINE_SIZE, 256 socket_id); 257 if (!txq) 258 return -ENOMEM; 259 260 txq->hv = hv; 261 txq->chan = hv->channels[queue_idx]; 262 txq->port_id = dev->data->port_id; 263 txq->queue_id = queue_idx; 264 265 tx_free_thresh = tx_conf->tx_free_thresh; 266 if (tx_free_thresh == 0) 267 tx_free_thresh = RTE_MIN(nb_desc / 4, 268 DEFAULT_TX_FREE_THRESH); 269 270 if (tx_free_thresh + 3 >= nb_desc) { 271 PMD_INIT_LOG(ERR, 272 "tx_free_thresh must be less than the number of TX entries minus 3(%u)." 273 " (tx_free_thresh=%u port=%u queue=%u)\n", 274 nb_desc - 3, 275 tx_free_thresh, dev->data->port_id, queue_idx); 276 return -EINVAL; 277 } 278 279 txq->free_thresh = tx_free_thresh; 280 281 snprintf(name, sizeof(name), 282 "hn_txd_%u_%u", dev->data->port_id, queue_idx); 283 284 PMD_INIT_LOG(DEBUG, "TX descriptor pool %s n=%u size=%zu", 285 name, nb_desc, sizeof(struct hn_txdesc)); 286 287 txq->tx_rndis = rte_calloc("hn_txq_rndis", nb_desc, 288 HN_RNDIS_PKT_ALIGNED, RTE_CACHE_LINE_SIZE); 289 if (txq->tx_rndis == NULL) 290 goto error; 291 292 txq->txdesc_pool = rte_mempool_create(name, nb_desc, 293 sizeof(struct hn_txdesc), 294 0, 0, NULL, NULL, 295 hn_txd_init, txq, 296 dev->device->numa_node, 0); 297 if (txq->txdesc_pool == NULL) { 298 PMD_DRV_LOG(ERR, 299 "mempool %s create failed: %d", name, rte_errno); 300 goto error; 301 } 302 303 txq->agg_szmax = RTE_MIN(hv->chim_szmax, hv->rndis_agg_size); 304 txq->agg_pktmax = hv->rndis_agg_pkts; 305 txq->agg_align = hv->rndis_agg_align; 306 307 hn_reset_txagg(txq); 308 309 err = hn_vf_tx_queue_setup(dev, queue_idx, nb_desc, 310 socket_id, tx_conf); 311 if (err == 0) { 312 dev->data->tx_queues[queue_idx] = txq; 313 return 0; 314 } 315 316 error: 317 if (txq->txdesc_pool) 318 rte_mempool_free(txq->txdesc_pool); 319 rte_free(txq->tx_rndis); 320 rte_free(txq); 321 return err; 322 } 323 324 void 325 hn_dev_tx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id, 326 struct rte_eth_txq_info *qinfo) 327 { 328 struct hn_tx_queue *txq = dev->data->tx_queues[queue_id]; 329 330 qinfo->nb_desc = txq->txdesc_pool->size; 331 qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads; 332 } 333 334 static struct hn_txdesc *hn_txd_get(struct hn_tx_queue *txq) 335 { 336 struct hn_txdesc *txd; 337 338 if (rte_mempool_get(txq->txdesc_pool, (void **)&txd)) { 339 ++txq->stats.ring_full; 340 PMD_TX_LOG(DEBUG, "tx pool exhausted!"); 341 return NULL; 342 } 343 344 txd->m = NULL; 345 txd->packets = 0; 346 txd->data_size = 0; 347 txd->chim_size = 0; 348 349 return txd; 350 } 351 352 static void hn_txd_put(struct hn_tx_queue *txq, struct hn_txdesc *txd) 353 { 354 rte_mempool_put(txq->txdesc_pool, txd); 355 } 356 357 void 358 hn_dev_tx_queue_release(void *arg) 359 { 360 struct hn_tx_queue *txq = arg; 361 362 PMD_INIT_FUNC_TRACE(); 363 364 if (!txq) 365 return; 366 367 if (txq->txdesc_pool) 368 rte_mempool_free(txq->txdesc_pool); 369 370 rte_free(txq->tx_rndis); 371 rte_free(txq); 372 } 373 374 /* 375 * Check the status of a Tx descriptor in the queue. 376 * 377 * returns: 378 * - -EINVAL - offset outside of tx_descriptor pool. 379 * - RTE_ETH_TX_DESC_FULL - descriptor is not acknowledged by host. 380 * - RTE_ETH_TX_DESC_DONE - descriptor is available. 381 */ 382 int hn_dev_tx_descriptor_status(void *arg, uint16_t offset) 383 { 384 const struct hn_tx_queue *txq = arg; 385 386 hn_process_events(txq->hv, txq->queue_id, 0); 387 388 if (offset >= rte_mempool_avail_count(txq->txdesc_pool)) 389 return -EINVAL; 390 391 if (offset < rte_mempool_in_use_count(txq->txdesc_pool)) 392 return RTE_ETH_TX_DESC_FULL; 393 else 394 return RTE_ETH_TX_DESC_DONE; 395 } 396 397 static void 398 hn_nvs_send_completed(struct rte_eth_dev *dev, uint16_t queue_id, 399 unsigned long xactid, const struct hn_nvs_rndis_ack *ack) 400 { 401 struct hn_data *hv = dev->data->dev_private; 402 struct hn_txdesc *txd = (struct hn_txdesc *)xactid; 403 struct hn_tx_queue *txq; 404 405 /* Control packets are sent with xacid == 0 */ 406 if (!txd) 407 return; 408 409 txq = dev->data->tx_queues[queue_id]; 410 if (likely(ack->status == NVS_STATUS_OK)) { 411 PMD_TX_LOG(DEBUG, "port %u:%u complete tx %u packets %u bytes %u", 412 txq->port_id, txq->queue_id, txd->chim_index, 413 txd->packets, txd->data_size); 414 txq->stats.bytes += txd->data_size; 415 txq->stats.packets += txd->packets; 416 } else { 417 PMD_DRV_LOG(NOTICE, "port %u:%u complete tx %u failed status %u", 418 txq->port_id, txq->queue_id, txd->chim_index, ack->status); 419 ++txq->stats.errors; 420 } 421 422 if (txd->chim_index != NVS_CHIM_IDX_INVALID) { 423 hn_chim_free(hv, txd->chim_index); 424 txd->chim_index = NVS_CHIM_IDX_INVALID; 425 } 426 427 rte_pktmbuf_free(txd->m); 428 hn_txd_put(txq, txd); 429 } 430 431 /* Handle transmit completion events */ 432 static void 433 hn_nvs_handle_comp(struct rte_eth_dev *dev, uint16_t queue_id, 434 const struct vmbus_chanpkt_hdr *pkt, 435 const void *data) 436 { 437 const struct hn_nvs_hdr *hdr = data; 438 439 switch (hdr->type) { 440 case NVS_TYPE_RNDIS_ACK: 441 hn_nvs_send_completed(dev, queue_id, pkt->xactid, data); 442 break; 443 444 default: 445 PMD_DRV_LOG(NOTICE, "unexpected send completion type %u", 446 hdr->type); 447 } 448 } 449 450 /* Parse per-packet info (meta data) */ 451 static int 452 hn_rndis_rxinfo(const void *info_data, unsigned int info_dlen, 453 struct hn_rxinfo *info) 454 { 455 const struct rndis_pktinfo *pi = info_data; 456 uint32_t mask = 0; 457 458 while (info_dlen != 0) { 459 const void *data; 460 uint32_t dlen; 461 462 if (unlikely(info_dlen < sizeof(*pi))) 463 return -EINVAL; 464 465 if (unlikely(info_dlen < pi->size)) 466 return -EINVAL; 467 info_dlen -= pi->size; 468 469 if (unlikely(pi->size & RNDIS_PKTINFO_SIZE_ALIGNMASK)) 470 return -EINVAL; 471 if (unlikely(pi->size < pi->offset)) 472 return -EINVAL; 473 474 dlen = pi->size - pi->offset; 475 data = pi->data; 476 477 switch (pi->type) { 478 case NDIS_PKTINFO_TYPE_VLAN: 479 if (unlikely(dlen < NDIS_VLAN_INFO_SIZE)) 480 return -EINVAL; 481 info->vlan_info = *((const uint32_t *)data); 482 mask |= HN_RXINFO_VLAN; 483 break; 484 485 case NDIS_PKTINFO_TYPE_CSUM: 486 if (unlikely(dlen < NDIS_RXCSUM_INFO_SIZE)) 487 return -EINVAL; 488 info->csum_info = *((const uint32_t *)data); 489 mask |= HN_RXINFO_CSUM; 490 break; 491 492 case NDIS_PKTINFO_TYPE_HASHVAL: 493 if (unlikely(dlen < NDIS_HASH_VALUE_SIZE)) 494 return -EINVAL; 495 info->hash_value = *((const uint32_t *)data); 496 mask |= HN_RXINFO_HASHVAL; 497 break; 498 499 case NDIS_PKTINFO_TYPE_HASHINF: 500 if (unlikely(dlen < NDIS_HASH_INFO_SIZE)) 501 return -EINVAL; 502 info->hash_info = *((const uint32_t *)data); 503 mask |= HN_RXINFO_HASHINF; 504 break; 505 506 default: 507 goto next; 508 } 509 510 if (mask == HN_RXINFO_ALL) 511 break; /* All found; done */ 512 next: 513 pi = (const struct rndis_pktinfo *) 514 ((const uint8_t *)pi + pi->size); 515 } 516 517 /* 518 * Final fixup. 519 * - If there is no hash value, invalidate the hash info. 520 */ 521 if (!(mask & HN_RXINFO_HASHVAL)) 522 info->hash_info = HN_NDIS_HASH_INFO_INVALID; 523 return 0; 524 } 525 526 static void hn_rx_buf_free_cb(void *buf __rte_unused, void *opaque) 527 { 528 struct hn_rx_bufinfo *rxb = opaque; 529 struct hn_rx_queue *rxq = rxb->rxq; 530 531 rte_atomic32_dec(&rxq->rxbuf_outstanding); 532 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid); 533 } 534 535 static struct hn_rx_bufinfo *hn_rx_buf_init(struct hn_rx_queue *rxq, 536 const struct vmbus_chanpkt_rxbuf *pkt) 537 { 538 struct hn_rx_bufinfo *rxb; 539 540 rxb = rxq->rxbuf_info + pkt->hdr.xactid; 541 rxb->chan = rxq->chan; 542 rxb->xactid = pkt->hdr.xactid; 543 rxb->rxq = rxq; 544 545 rxb->shinfo.free_cb = hn_rx_buf_free_cb; 546 rxb->shinfo.fcb_opaque = rxb; 547 rte_mbuf_ext_refcnt_set(&rxb->shinfo, 1); 548 return rxb; 549 } 550 551 static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb, 552 uint8_t *data, unsigned int headroom, unsigned int dlen, 553 const struct hn_rxinfo *info) 554 { 555 struct hn_data *hv = rxq->hv; 556 struct rte_mbuf *m; 557 bool use_extbuf = false; 558 559 m = rte_pktmbuf_alloc(rxq->mb_pool); 560 if (unlikely(!m)) { 561 struct rte_eth_dev *dev = 562 &rte_eth_devices[rxq->port_id]; 563 564 dev->data->rx_mbuf_alloc_failed++; 565 return; 566 } 567 568 /* 569 * For large packets, avoid copy if possible but need to keep 570 * some space available in receive area for later packets. 571 */ 572 if (dlen >= HN_RXCOPY_THRESHOLD && 573 (uint32_t)rte_atomic32_read(&rxq->rxbuf_outstanding) < 574 hv->rxbuf_section_cnt / 2) { 575 struct rte_mbuf_ext_shared_info *shinfo; 576 const void *rxbuf; 577 rte_iova_t iova; 578 579 /* 580 * Build an external mbuf that points to recveive area. 581 * Use refcount to handle multiple packets in same 582 * receive buffer section. 583 */ 584 rxbuf = hv->rxbuf_res->addr; 585 iova = rte_mem_virt2iova(rxbuf) + RTE_PTR_DIFF(data, rxbuf); 586 shinfo = &rxb->shinfo; 587 588 /* shinfo is already set to 1 by the caller */ 589 if (rte_mbuf_ext_refcnt_update(shinfo, 1) == 2) 590 rte_atomic32_inc(&rxq->rxbuf_outstanding); 591 592 rte_pktmbuf_attach_extbuf(m, data, iova, 593 dlen + headroom, shinfo); 594 m->data_off = headroom; 595 use_extbuf = true; 596 } else { 597 /* Mbuf's in pool must be large enough to hold small packets */ 598 if (unlikely(rte_pktmbuf_tailroom(m) < dlen)) { 599 rte_pktmbuf_free_seg(m); 600 ++rxq->stats.errors; 601 return; 602 } 603 rte_memcpy(rte_pktmbuf_mtod(m, void *), 604 data + headroom, dlen); 605 } 606 607 m->port = rxq->port_id; 608 m->pkt_len = dlen; 609 m->data_len = dlen; 610 m->packet_type = rte_net_get_ptype(m, NULL, 611 RTE_PTYPE_L2_MASK | 612 RTE_PTYPE_L3_MASK | 613 RTE_PTYPE_L4_MASK); 614 615 if (info->vlan_info != HN_NDIS_VLAN_INFO_INVALID) { 616 m->vlan_tci = info->vlan_info; 617 m->ol_flags |= PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN; 618 619 /* NDIS always strips tag, put it back if necessary */ 620 if (!hv->vlan_strip && rte_vlan_insert(&m)) { 621 PMD_DRV_LOG(DEBUG, "vlan insert failed"); 622 ++rxq->stats.errors; 623 if (use_extbuf) 624 rte_pktmbuf_detach_extbuf(m); 625 rte_pktmbuf_free(m); 626 return; 627 } 628 } 629 630 if (info->csum_info != HN_NDIS_RXCSUM_INFO_INVALID) { 631 if (info->csum_info & NDIS_RXCSUM_INFO_IPCS_OK) 632 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD; 633 634 if (info->csum_info & (NDIS_RXCSUM_INFO_UDPCS_OK 635 | NDIS_RXCSUM_INFO_TCPCS_OK)) 636 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD; 637 else if (info->csum_info & (NDIS_RXCSUM_INFO_TCPCS_FAILED 638 | NDIS_RXCSUM_INFO_UDPCS_FAILED)) 639 m->ol_flags |= PKT_RX_L4_CKSUM_BAD; 640 } 641 642 if (info->hash_info != HN_NDIS_HASH_INFO_INVALID) { 643 m->ol_flags |= PKT_RX_RSS_HASH; 644 m->hash.rss = info->hash_value; 645 } 646 647 PMD_RX_LOG(DEBUG, 648 "port %u:%u RX id %"PRIu64" size %u type %#x ol_flags %#"PRIx64, 649 rxq->port_id, rxq->queue_id, rxb->xactid, 650 m->pkt_len, m->packet_type, m->ol_flags); 651 652 ++rxq->stats.packets; 653 rxq->stats.bytes += m->pkt_len; 654 hn_update_packet_stats(&rxq->stats, m); 655 656 if (unlikely(rte_ring_sp_enqueue(rxq->rx_ring, m) != 0)) { 657 ++rxq->stats.ring_full; 658 PMD_RX_LOG(DEBUG, "rx ring full"); 659 if (use_extbuf) 660 rte_pktmbuf_detach_extbuf(m); 661 rte_pktmbuf_free(m); 662 } 663 } 664 665 static void hn_rndis_rx_data(struct hn_rx_queue *rxq, 666 struct hn_rx_bufinfo *rxb, 667 void *data, uint32_t dlen) 668 { 669 unsigned int data_off, data_len; 670 unsigned int pktinfo_off, pktinfo_len; 671 const struct rndis_packet_msg *pkt = data; 672 struct hn_rxinfo info = { 673 .vlan_info = HN_NDIS_VLAN_INFO_INVALID, 674 .csum_info = HN_NDIS_RXCSUM_INFO_INVALID, 675 .hash_info = HN_NDIS_HASH_INFO_INVALID, 676 }; 677 int err; 678 679 hn_rndis_dump(pkt); 680 681 if (unlikely(dlen < sizeof(*pkt))) 682 goto error; 683 684 if (unlikely(dlen < pkt->len)) 685 goto error; /* truncated RNDIS from host */ 686 687 if (unlikely(pkt->len < pkt->datalen 688 + pkt->oobdatalen + pkt->pktinfolen)) 689 goto error; 690 691 if (unlikely(pkt->datalen == 0)) 692 goto error; 693 694 /* Check offsets. */ 695 if (unlikely(pkt->dataoffset < RNDIS_PACKET_MSG_OFFSET_MIN)) 696 goto error; 697 698 if (likely(pkt->pktinfooffset > 0) && 699 unlikely(pkt->pktinfooffset < RNDIS_PACKET_MSG_OFFSET_MIN || 700 (pkt->pktinfooffset & RNDIS_PACKET_MSG_OFFSET_ALIGNMASK))) 701 goto error; 702 703 data_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset); 704 data_len = pkt->datalen; 705 pktinfo_off = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->pktinfooffset); 706 pktinfo_len = pkt->pktinfolen; 707 708 if (likely(pktinfo_len > 0)) { 709 err = hn_rndis_rxinfo((const uint8_t *)pkt + pktinfo_off, 710 pktinfo_len, &info); 711 if (err) 712 goto error; 713 } 714 715 /* overflow check */ 716 if (data_len > data_len + data_off || data_len + data_off > pkt->len) 717 goto error; 718 719 if (unlikely(data_len < RTE_ETHER_HDR_LEN)) 720 goto error; 721 722 hn_rxpkt(rxq, rxb, data, data_off, data_len, &info); 723 return; 724 error: 725 ++rxq->stats.errors; 726 } 727 728 static void 729 hn_rndis_receive(struct rte_eth_dev *dev, struct hn_rx_queue *rxq, 730 struct hn_rx_bufinfo *rxb, void *buf, uint32_t len) 731 { 732 const struct rndis_msghdr *hdr = buf; 733 734 switch (hdr->type) { 735 case RNDIS_PACKET_MSG: 736 if (dev->data->dev_started) 737 hn_rndis_rx_data(rxq, rxb, buf, len); 738 break; 739 740 case RNDIS_INDICATE_STATUS_MSG: 741 hn_rndis_link_status(dev, buf); 742 break; 743 744 case RNDIS_INITIALIZE_CMPLT: 745 case RNDIS_QUERY_CMPLT: 746 case RNDIS_SET_CMPLT: 747 hn_rndis_receive_response(rxq->hv, buf, len); 748 break; 749 750 default: 751 PMD_DRV_LOG(NOTICE, 752 "unexpected RNDIS message (type %#x len %u)", 753 hdr->type, len); 754 break; 755 } 756 } 757 758 static void 759 hn_nvs_handle_rxbuf(struct rte_eth_dev *dev, 760 struct hn_data *hv, 761 struct hn_rx_queue *rxq, 762 const struct vmbus_chanpkt_hdr *hdr, 763 const void *buf) 764 { 765 const struct vmbus_chanpkt_rxbuf *pkt; 766 const struct hn_nvs_hdr *nvs_hdr = buf; 767 uint32_t rxbuf_sz = hv->rxbuf_res->len; 768 char *rxbuf = hv->rxbuf_res->addr; 769 unsigned int i, hlen, count; 770 struct hn_rx_bufinfo *rxb; 771 772 /* At minimum we need type header */ 773 if (unlikely(vmbus_chanpkt_datalen(hdr) < sizeof(*nvs_hdr))) { 774 PMD_RX_LOG(ERR, "invalid receive nvs RNDIS"); 775 return; 776 } 777 778 /* Make sure that this is a RNDIS message. */ 779 if (unlikely(nvs_hdr->type != NVS_TYPE_RNDIS)) { 780 PMD_RX_LOG(ERR, "nvs type %u, not RNDIS", 781 nvs_hdr->type); 782 return; 783 } 784 785 hlen = vmbus_chanpkt_getlen(hdr->hlen); 786 if (unlikely(hlen < sizeof(*pkt))) { 787 PMD_RX_LOG(ERR, "invalid rxbuf chanpkt"); 788 return; 789 } 790 791 pkt = container_of(hdr, const struct vmbus_chanpkt_rxbuf, hdr); 792 if (unlikely(pkt->rxbuf_id != NVS_RXBUF_SIG)) { 793 PMD_RX_LOG(ERR, "invalid rxbuf_id 0x%08x", 794 pkt->rxbuf_id); 795 return; 796 } 797 798 count = pkt->rxbuf_cnt; 799 if (unlikely(hlen < offsetof(struct vmbus_chanpkt_rxbuf, 800 rxbuf[count]))) { 801 PMD_RX_LOG(ERR, "invalid rxbuf_cnt %u", count); 802 return; 803 } 804 805 if (pkt->hdr.xactid > hv->rxbuf_section_cnt) { 806 PMD_RX_LOG(ERR, "invalid rxbuf section id %" PRIx64, 807 pkt->hdr.xactid); 808 return; 809 } 810 811 /* Setup receive buffer info to allow for callback */ 812 rxb = hn_rx_buf_init(rxq, pkt); 813 814 /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */ 815 for (i = 0; i < count; ++i) { 816 unsigned int ofs, len; 817 818 ofs = pkt->rxbuf[i].ofs; 819 len = pkt->rxbuf[i].len; 820 821 if (unlikely(ofs + len > rxbuf_sz)) { 822 PMD_RX_LOG(ERR, 823 "%uth RNDIS msg overflow ofs %u, len %u", 824 i, ofs, len); 825 continue; 826 } 827 828 if (unlikely(len == 0)) { 829 PMD_RX_LOG(ERR, "%uth RNDIS msg len %u", i, len); 830 continue; 831 } 832 833 hn_rndis_receive(dev, rxq, rxb, 834 rxbuf + ofs, len); 835 } 836 837 /* Send ACK now if external mbuf not used */ 838 if (rte_mbuf_ext_refcnt_update(&rxb->shinfo, -1) == 0) 839 hn_nvs_ack_rxbuf(rxb->chan, rxb->xactid); 840 } 841 842 /* 843 * Called when NVS inband events are received. 844 * Send up a two part message with port_id and the NVS message 845 * to the pipe to the netvsc-vf-event control thread. 846 */ 847 static void hn_nvs_handle_notify(struct rte_eth_dev *dev, 848 const struct vmbus_chanpkt_hdr *pkt, 849 const void *data) 850 { 851 const struct hn_nvs_hdr *hdr = data; 852 853 switch (hdr->type) { 854 case NVS_TYPE_TXTBL_NOTE: 855 /* Transmit indirection table has locking problems 856 * in DPDK and therefore not implemented 857 */ 858 PMD_DRV_LOG(DEBUG, "host notify of transmit indirection table"); 859 break; 860 861 case NVS_TYPE_VFASSOC_NOTE: 862 hn_nvs_handle_vfassoc(dev, pkt, data); 863 break; 864 865 default: 866 PMD_DRV_LOG(INFO, 867 "got notify, nvs type %u", hdr->type); 868 } 869 } 870 871 struct hn_rx_queue *hn_rx_queue_alloc(struct hn_data *hv, 872 uint16_t queue_id, 873 unsigned int socket_id) 874 { 875 struct hn_rx_queue *rxq; 876 877 rxq = rte_zmalloc_socket("HN_RXQ", sizeof(*rxq), 878 RTE_CACHE_LINE_SIZE, socket_id); 879 if (!rxq) 880 return NULL; 881 882 rxq->hv = hv; 883 rxq->chan = hv->channels[queue_id]; 884 rte_spinlock_init(&rxq->ring_lock); 885 rxq->port_id = hv->port_id; 886 rxq->queue_id = queue_id; 887 rxq->event_sz = HN_RXQ_EVENT_DEFAULT; 888 rxq->event_buf = rte_malloc_socket("HN_EVENTS", HN_RXQ_EVENT_DEFAULT, 889 RTE_CACHE_LINE_SIZE, socket_id); 890 if (!rxq->event_buf) { 891 rte_free(rxq); 892 return NULL; 893 } 894 895 /* setup rxbuf_info for non-primary queue */ 896 if (queue_id) { 897 rxq->rxbuf_info = rte_calloc("HN_RXBUF_INFO", 898 hv->rxbuf_section_cnt, 899 sizeof(*rxq->rxbuf_info), 900 RTE_CACHE_LINE_SIZE); 901 902 if (!rxq->rxbuf_info) { 903 PMD_DRV_LOG(ERR, 904 "Could not allocate rxbuf info for queue %d\n", 905 queue_id); 906 rte_free(rxq->event_buf); 907 rte_free(rxq); 908 return NULL; 909 } 910 } 911 912 return rxq; 913 } 914 915 void 916 hn_dev_rx_queue_info(struct rte_eth_dev *dev, uint16_t queue_id, 917 struct rte_eth_rxq_info *qinfo) 918 { 919 struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id]; 920 921 qinfo->mp = rxq->mb_pool; 922 qinfo->nb_desc = rxq->rx_ring->size; 923 qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads; 924 } 925 926 int 927 hn_dev_rx_queue_setup(struct rte_eth_dev *dev, 928 uint16_t queue_idx, uint16_t nb_desc, 929 unsigned int socket_id, 930 const struct rte_eth_rxconf *rx_conf, 931 struct rte_mempool *mp) 932 { 933 struct hn_data *hv = dev->data->dev_private; 934 char ring_name[RTE_RING_NAMESIZE]; 935 struct hn_rx_queue *rxq; 936 unsigned int count; 937 int error = -ENOMEM; 938 939 PMD_INIT_FUNC_TRACE(); 940 941 if (queue_idx == 0) { 942 rxq = hv->primary; 943 } else { 944 rxq = hn_rx_queue_alloc(hv, queue_idx, socket_id); 945 if (!rxq) 946 return -ENOMEM; 947 } 948 949 rxq->mb_pool = mp; 950 count = rte_mempool_avail_count(mp) / dev->data->nb_rx_queues; 951 if (nb_desc == 0 || nb_desc > count) 952 nb_desc = count; 953 954 /* 955 * Staging ring from receive event logic to rx_pkts. 956 * rx_pkts assumes caller is handling multi-thread issue. 957 * event logic has locking. 958 */ 959 snprintf(ring_name, sizeof(ring_name), 960 "hn_rx_%u_%u", dev->data->port_id, queue_idx); 961 rxq->rx_ring = rte_ring_create(ring_name, 962 rte_align32pow2(nb_desc), 963 socket_id, 0); 964 if (!rxq->rx_ring) 965 goto fail; 966 967 error = hn_vf_rx_queue_setup(dev, queue_idx, nb_desc, 968 socket_id, rx_conf, mp); 969 if (error) 970 goto fail; 971 972 dev->data->rx_queues[queue_idx] = rxq; 973 return 0; 974 975 fail: 976 rte_ring_free(rxq->rx_ring); 977 rte_free(rxq->rxbuf_info); 978 rte_free(rxq->event_buf); 979 rte_free(rxq); 980 return error; 981 } 982 983 static void 984 hn_rx_queue_free(struct hn_rx_queue *rxq, bool keep_primary) 985 { 986 987 if (!rxq) 988 return; 989 990 rte_ring_free(rxq->rx_ring); 991 rxq->rx_ring = NULL; 992 rxq->mb_pool = NULL; 993 994 hn_vf_rx_queue_release(rxq->hv, rxq->queue_id); 995 996 /* Keep primary queue to allow for control operations */ 997 if (keep_primary && rxq == rxq->hv->primary) 998 return; 999 1000 rte_free(rxq->rxbuf_info); 1001 rte_free(rxq->event_buf); 1002 rte_free(rxq); 1003 } 1004 1005 void 1006 hn_dev_rx_queue_release(void *arg) 1007 { 1008 struct hn_rx_queue *rxq = arg; 1009 1010 PMD_INIT_FUNC_TRACE(); 1011 1012 hn_rx_queue_free(rxq, true); 1013 } 1014 1015 /* 1016 * Get the number of used descriptor in a rx queue 1017 * For this device that means how many packets are pending in the ring. 1018 */ 1019 uint32_t 1020 hn_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_id) 1021 { 1022 struct hn_rx_queue *rxq = dev->data->rx_queues[queue_id]; 1023 1024 return rte_ring_count(rxq->rx_ring); 1025 } 1026 1027 /* 1028 * Check the status of a Rx descriptor in the queue 1029 * 1030 * returns: 1031 * - -EINVAL - offset outside of ring 1032 * - RTE_ETH_RX_DESC_AVAIL - no data available yet 1033 * - RTE_ETH_RX_DESC_DONE - data is waiting in stagin ring 1034 */ 1035 int hn_dev_rx_queue_status(void *arg, uint16_t offset) 1036 { 1037 const struct hn_rx_queue *rxq = arg; 1038 1039 hn_process_events(rxq->hv, rxq->queue_id, 0); 1040 if (offset >= rxq->rx_ring->capacity) 1041 return -EINVAL; 1042 1043 if (offset < rte_ring_count(rxq->rx_ring)) 1044 return RTE_ETH_RX_DESC_DONE; 1045 else 1046 return RTE_ETH_RX_DESC_AVAIL; 1047 } 1048 1049 int 1050 hn_dev_tx_done_cleanup(void *arg, uint32_t free_cnt) 1051 { 1052 struct hn_tx_queue *txq = arg; 1053 1054 return hn_process_events(txq->hv, txq->queue_id, free_cnt); 1055 } 1056 1057 /* 1058 * Process pending events on the channel. 1059 * Called from both Rx queue poll and Tx cleanup 1060 */ 1061 uint32_t hn_process_events(struct hn_data *hv, uint16_t queue_id, 1062 uint32_t tx_limit) 1063 { 1064 struct rte_eth_dev *dev = &rte_eth_devices[hv->port_id]; 1065 struct hn_rx_queue *rxq; 1066 uint32_t bytes_read = 0; 1067 uint32_t tx_done = 0; 1068 int ret = 0; 1069 1070 rxq = queue_id == 0 ? hv->primary : dev->data->rx_queues[queue_id]; 1071 1072 /* 1073 * Since channel is shared between Rx and TX queue need to have a lock 1074 * since DPDK does not force same CPU to be used for Rx/Tx. 1075 */ 1076 if (unlikely(!rte_spinlock_trylock(&rxq->ring_lock))) 1077 return 0; 1078 1079 for (;;) { 1080 const struct vmbus_chanpkt_hdr *pkt; 1081 uint32_t len = rxq->event_sz; 1082 const void *data; 1083 1084 retry: 1085 ret = rte_vmbus_chan_recv_raw(rxq->chan, rxq->event_buf, &len); 1086 if (ret == -EAGAIN) 1087 break; /* ring is empty */ 1088 1089 if (unlikely(ret == -ENOBUFS)) { 1090 /* event buffer not large enough to read ring */ 1091 1092 PMD_DRV_LOG(DEBUG, 1093 "event buffer expansion (need %u)", len); 1094 rxq->event_sz = len + len / 4; 1095 rxq->event_buf = rte_realloc(rxq->event_buf, rxq->event_sz, 1096 RTE_CACHE_LINE_SIZE); 1097 if (rxq->event_buf) 1098 goto retry; 1099 /* out of memory, no more events now */ 1100 rxq->event_sz = 0; 1101 break; 1102 } 1103 1104 if (unlikely(ret <= 0)) { 1105 /* This indicates a failure to communicate (or worse) */ 1106 rte_exit(EXIT_FAILURE, 1107 "vmbus ring buffer error: %d", ret); 1108 } 1109 1110 bytes_read += ret; 1111 pkt = (const struct vmbus_chanpkt_hdr *)rxq->event_buf; 1112 data = (char *)rxq->event_buf + vmbus_chanpkt_getlen(pkt->hlen); 1113 1114 switch (pkt->type) { 1115 case VMBUS_CHANPKT_TYPE_COMP: 1116 ++tx_done; 1117 hn_nvs_handle_comp(dev, queue_id, pkt, data); 1118 break; 1119 1120 case VMBUS_CHANPKT_TYPE_RXBUF: 1121 hn_nvs_handle_rxbuf(dev, hv, rxq, pkt, data); 1122 break; 1123 1124 case VMBUS_CHANPKT_TYPE_INBAND: 1125 hn_nvs_handle_notify(dev, pkt, data); 1126 break; 1127 1128 default: 1129 PMD_DRV_LOG(ERR, "unknown chan pkt %u", pkt->type); 1130 break; 1131 } 1132 1133 if (tx_limit && tx_done >= tx_limit) 1134 break; 1135 } 1136 1137 if (bytes_read > 0) 1138 rte_vmbus_chan_signal_read(rxq->chan, bytes_read); 1139 1140 rte_spinlock_unlock(&rxq->ring_lock); 1141 1142 return tx_done; 1143 } 1144 1145 static void hn_append_to_chim(struct hn_tx_queue *txq, 1146 struct rndis_packet_msg *pkt, 1147 const struct rte_mbuf *m) 1148 { 1149 struct hn_txdesc *txd = txq->agg_txd; 1150 uint8_t *buf = (uint8_t *)pkt; 1151 unsigned int data_offs; 1152 1153 hn_rndis_dump(pkt); 1154 1155 data_offs = RNDIS_PACKET_MSG_OFFSET_ABS(pkt->dataoffset); 1156 txd->chim_size += pkt->len; 1157 txd->data_size += m->pkt_len; 1158 ++txd->packets; 1159 hn_update_packet_stats(&txq->stats, m); 1160 1161 for (; m; m = m->next) { 1162 uint16_t len = rte_pktmbuf_data_len(m); 1163 1164 rte_memcpy(buf + data_offs, 1165 rte_pktmbuf_mtod(m, const char *), len); 1166 data_offs += len; 1167 } 1168 } 1169 1170 /* 1171 * Send pending aggregated data in chimney buffer (if any). 1172 * Returns error if send was unsuccessful because channel ring buffer 1173 * was full. 1174 */ 1175 static int hn_flush_txagg(struct hn_tx_queue *txq, bool *need_sig) 1176 1177 { 1178 struct hn_txdesc *txd = txq->agg_txd; 1179 struct hn_nvs_rndis rndis; 1180 int ret; 1181 1182 if (!txd) 1183 return 0; 1184 1185 rndis = (struct hn_nvs_rndis) { 1186 .type = NVS_TYPE_RNDIS, 1187 .rndis_mtype = NVS_RNDIS_MTYPE_DATA, 1188 .chim_idx = txd->chim_index, 1189 .chim_sz = txd->chim_size, 1190 }; 1191 1192 PMD_TX_LOG(DEBUG, "port %u:%u tx %u size %u", 1193 txq->port_id, txq->queue_id, txd->chim_index, txd->chim_size); 1194 1195 ret = hn_nvs_send(txq->chan, VMBUS_CHANPKT_FLAG_RC, 1196 &rndis, sizeof(rndis), (uintptr_t)txd, need_sig); 1197 1198 if (likely(ret == 0)) 1199 hn_reset_txagg(txq); 1200 else if (ret == -EAGAIN) { 1201 PMD_TX_LOG(DEBUG, "port %u:%u channel full", 1202 txq->port_id, txq->queue_id); 1203 ++txq->stats.channel_full; 1204 } else { 1205 ++txq->stats.errors; 1206 1207 PMD_DRV_LOG(NOTICE, "port %u:%u send failed: %d", 1208 txq->port_id, txq->queue_id, ret); 1209 } 1210 return ret; 1211 } 1212 1213 /* 1214 * Try and find a place in a send chimney buffer to put 1215 * the small packet. If space is available, this routine 1216 * returns a pointer of where to place the data. 1217 * If no space, caller should try direct transmit. 1218 */ 1219 static void * 1220 hn_try_txagg(struct hn_data *hv, struct hn_tx_queue *txq, 1221 struct hn_txdesc *txd, uint32_t pktsize) 1222 { 1223 struct hn_txdesc *agg_txd = txq->agg_txd; 1224 struct rndis_packet_msg *pkt; 1225 void *chim; 1226 1227 if (agg_txd) { 1228 unsigned int padding, olen; 1229 1230 /* 1231 * Update the previous RNDIS packet's total length, 1232 * it can be increased due to the mandatory alignment 1233 * padding for this RNDIS packet. And update the 1234 * aggregating txdesc's chimney sending buffer size 1235 * accordingly. 1236 * 1237 * Zero-out the padding, as required by the RNDIS spec. 1238 */ 1239 pkt = txq->agg_prevpkt; 1240 olen = pkt->len; 1241 padding = RTE_ALIGN(olen, txq->agg_align) - olen; 1242 if (padding > 0) { 1243 agg_txd->chim_size += padding; 1244 pkt->len += padding; 1245 memset((uint8_t *)pkt + olen, 0, padding); 1246 } 1247 1248 chim = (uint8_t *)pkt + pkt->len; 1249 txq->agg_prevpkt = chim; 1250 txq->agg_pktleft--; 1251 txq->agg_szleft -= pktsize; 1252 if (txq->agg_szleft < HN_PKTSIZE_MIN(txq->agg_align)) { 1253 /* 1254 * Probably can't aggregate more packets, 1255 * flush this aggregating txdesc proactively. 1256 */ 1257 txq->agg_pktleft = 0; 1258 } 1259 1260 hn_txd_put(txq, txd); 1261 return chim; 1262 } 1263 1264 txd->chim_index = hn_chim_alloc(hv); 1265 if (txd->chim_index == NVS_CHIM_IDX_INVALID) 1266 return NULL; 1267 1268 chim = (uint8_t *)hv->chim_res->addr 1269 + txd->chim_index * hv->chim_szmax; 1270 1271 txq->agg_txd = txd; 1272 txq->agg_pktleft = txq->agg_pktmax - 1; 1273 txq->agg_szleft = txq->agg_szmax - pktsize; 1274 txq->agg_prevpkt = chim; 1275 1276 return chim; 1277 } 1278 1279 static inline void * 1280 hn_rndis_pktinfo_append(struct rndis_packet_msg *pkt, 1281 uint32_t pi_dlen, uint32_t pi_type) 1282 { 1283 const uint32_t pi_size = RNDIS_PKTINFO_SIZE(pi_dlen); 1284 struct rndis_pktinfo *pi; 1285 1286 /* 1287 * Per-packet-info does not move; it only grows. 1288 * 1289 * NOTE: 1290 * pktinfooffset in this phase counts from the beginning 1291 * of rndis_packet_msg. 1292 */ 1293 pi = (struct rndis_pktinfo *)((uint8_t *)pkt + hn_rndis_pktlen(pkt)); 1294 1295 pkt->pktinfolen += pi_size; 1296 1297 pi->size = pi_size; 1298 pi->type = pi_type; 1299 pi->offset = RNDIS_PKTINFO_OFFSET; 1300 1301 return pi->data; 1302 } 1303 1304 /* Put RNDIS header and packet info on packet */ 1305 static void hn_encap(struct rndis_packet_msg *pkt, 1306 uint16_t queue_id, 1307 const struct rte_mbuf *m) 1308 { 1309 unsigned int hlen = m->l2_len + m->l3_len; 1310 uint32_t *pi_data; 1311 uint32_t pkt_hlen; 1312 1313 pkt->type = RNDIS_PACKET_MSG; 1314 pkt->len = m->pkt_len; 1315 pkt->dataoffset = 0; 1316 pkt->datalen = m->pkt_len; 1317 pkt->oobdataoffset = 0; 1318 pkt->oobdatalen = 0; 1319 pkt->oobdataelements = 0; 1320 pkt->pktinfooffset = sizeof(*pkt); 1321 pkt->pktinfolen = 0; 1322 pkt->vchandle = 0; 1323 pkt->reserved = 0; 1324 1325 /* 1326 * Set the hash value for this packet, to the queue_id to cause 1327 * TX done event for this packet on the right channel. 1328 */ 1329 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_HASH_VALUE_SIZE, 1330 NDIS_PKTINFO_TYPE_HASHVAL); 1331 *pi_data = queue_id; 1332 1333 if (m->ol_flags & PKT_TX_VLAN_PKT) { 1334 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_VLAN_INFO_SIZE, 1335 NDIS_PKTINFO_TYPE_VLAN); 1336 *pi_data = m->vlan_tci; 1337 } 1338 1339 if (m->ol_flags & PKT_TX_TCP_SEG) { 1340 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_LSO2_INFO_SIZE, 1341 NDIS_PKTINFO_TYPE_LSO); 1342 1343 if (m->ol_flags & PKT_TX_IPV6) { 1344 *pi_data = NDIS_LSO2_INFO_MAKEIPV6(hlen, 1345 m->tso_segsz); 1346 } else { 1347 *pi_data = NDIS_LSO2_INFO_MAKEIPV4(hlen, 1348 m->tso_segsz); 1349 } 1350 } else if (m->ol_flags & 1351 (PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM | PKT_TX_IP_CKSUM)) { 1352 pi_data = hn_rndis_pktinfo_append(pkt, NDIS_TXCSUM_INFO_SIZE, 1353 NDIS_PKTINFO_TYPE_CSUM); 1354 *pi_data = 0; 1355 1356 if (m->ol_flags & PKT_TX_IPV6) 1357 *pi_data |= NDIS_TXCSUM_INFO_IPV6; 1358 if (m->ol_flags & PKT_TX_IPV4) { 1359 *pi_data |= NDIS_TXCSUM_INFO_IPV4; 1360 1361 if (m->ol_flags & PKT_TX_IP_CKSUM) 1362 *pi_data |= NDIS_TXCSUM_INFO_IPCS; 1363 } 1364 1365 if (m->ol_flags & PKT_TX_TCP_CKSUM) 1366 *pi_data |= NDIS_TXCSUM_INFO_MKTCPCS(hlen); 1367 else if (m->ol_flags & PKT_TX_UDP_CKSUM) 1368 *pi_data |= NDIS_TXCSUM_INFO_MKUDPCS(hlen); 1369 } 1370 1371 pkt_hlen = pkt->pktinfooffset + pkt->pktinfolen; 1372 /* Fixup RNDIS packet message total length */ 1373 pkt->len += pkt_hlen; 1374 1375 /* Convert RNDIS packet message offsets */ 1376 pkt->dataoffset = hn_rndis_pktmsg_offset(pkt_hlen); 1377 pkt->pktinfooffset = hn_rndis_pktmsg_offset(pkt->pktinfooffset); 1378 } 1379 1380 /* How many scatter gather list elements ar needed */ 1381 static unsigned int hn_get_slots(const struct rte_mbuf *m) 1382 { 1383 unsigned int slots = 1; /* for RNDIS header */ 1384 1385 while (m) { 1386 unsigned int size = rte_pktmbuf_data_len(m); 1387 unsigned int offs = rte_mbuf_data_iova(m) & PAGE_MASK; 1388 1389 slots += (offs + size + PAGE_SIZE - 1) / PAGE_SIZE; 1390 m = m->next; 1391 } 1392 1393 return slots; 1394 } 1395 1396 /* Build scatter gather list from chained mbuf */ 1397 static unsigned int hn_fill_sg(struct vmbus_gpa *sg, 1398 const struct rte_mbuf *m) 1399 { 1400 unsigned int segs = 0; 1401 1402 while (m) { 1403 rte_iova_t addr = rte_mbuf_data_iova(m); 1404 unsigned int page = addr / PAGE_SIZE; 1405 unsigned int offset = addr & PAGE_MASK; 1406 unsigned int len = rte_pktmbuf_data_len(m); 1407 1408 while (len > 0) { 1409 unsigned int bytes = RTE_MIN(len, PAGE_SIZE - offset); 1410 1411 sg[segs].page = page; 1412 sg[segs].ofs = offset; 1413 sg[segs].len = bytes; 1414 segs++; 1415 1416 ++page; 1417 offset = 0; 1418 len -= bytes; 1419 } 1420 m = m->next; 1421 } 1422 1423 return segs; 1424 } 1425 1426 /* Transmit directly from mbuf */ 1427 static int hn_xmit_sg(struct hn_tx_queue *txq, 1428 const struct hn_txdesc *txd, const struct rte_mbuf *m, 1429 bool *need_sig) 1430 { 1431 struct vmbus_gpa sg[hn_get_slots(m)]; 1432 struct hn_nvs_rndis nvs_rndis = { 1433 .type = NVS_TYPE_RNDIS, 1434 .rndis_mtype = NVS_RNDIS_MTYPE_DATA, 1435 .chim_sz = txd->chim_size, 1436 }; 1437 rte_iova_t addr; 1438 unsigned int segs; 1439 1440 /* attach aggregation data if present */ 1441 if (txd->chim_size > 0) 1442 nvs_rndis.chim_idx = txd->chim_index; 1443 else 1444 nvs_rndis.chim_idx = NVS_CHIM_IDX_INVALID; 1445 1446 hn_rndis_dump(txd->rndis_pkt); 1447 1448 /* pass IOVA of rndis header in first segment */ 1449 addr = rte_malloc_virt2iova(txq->tx_rndis); 1450 if (unlikely(addr == RTE_BAD_IOVA)) { 1451 PMD_DRV_LOG(ERR, "RNDIS transmit can not get iova"); 1452 return -EINVAL; 1453 } 1454 addr = addr + ((char *)txd->rndis_pkt - (char *)txq->tx_rndis); 1455 1456 sg[0].page = addr / PAGE_SIZE; 1457 sg[0].ofs = addr & PAGE_MASK; 1458 sg[0].len = RNDIS_PACKET_MSG_OFFSET_ABS(hn_rndis_pktlen(txd->rndis_pkt)); 1459 segs = 1; 1460 1461 hn_update_packet_stats(&txq->stats, m); 1462 1463 segs += hn_fill_sg(sg + 1, m); 1464 1465 PMD_TX_LOG(DEBUG, "port %u:%u tx %u segs %u size %u", 1466 txq->port_id, txq->queue_id, txd->chim_index, 1467 segs, nvs_rndis.chim_sz); 1468 1469 return hn_nvs_send_sglist(txq->chan, sg, segs, 1470 &nvs_rndis, sizeof(nvs_rndis), 1471 (uintptr_t)txd, need_sig); 1472 } 1473 1474 uint16_t 1475 hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts) 1476 { 1477 struct hn_tx_queue *txq = ptxq; 1478 uint16_t queue_id = txq->queue_id; 1479 struct hn_data *hv = txq->hv; 1480 struct rte_eth_dev *vf_dev; 1481 bool need_sig = false; 1482 uint16_t nb_tx, tx_thresh; 1483 int ret; 1484 1485 if (unlikely(hv->closed)) 1486 return 0; 1487 1488 /* 1489 * Always check for events on the primary channel 1490 * because that is where hotplug notifications occur. 1491 */ 1492 tx_thresh = RTE_MAX(txq->free_thresh, nb_pkts); 1493 if (txq->queue_id == 0 || 1494 rte_mempool_avail_count(txq->txdesc_pool) < tx_thresh) 1495 hn_process_events(hv, txq->queue_id, 0); 1496 1497 /* Transmit over VF if present and up */ 1498 rte_rwlock_read_lock(&hv->vf_lock); 1499 vf_dev = hn_get_vf_dev(hv); 1500 if (vf_dev && vf_dev->data->dev_started) { 1501 void *sub_q = vf_dev->data->tx_queues[queue_id]; 1502 1503 nb_tx = (*vf_dev->tx_pkt_burst)(sub_q, tx_pkts, nb_pkts); 1504 rte_rwlock_read_unlock(&hv->vf_lock); 1505 return nb_tx; 1506 } 1507 rte_rwlock_read_unlock(&hv->vf_lock); 1508 1509 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) { 1510 struct rte_mbuf *m = tx_pkts[nb_tx]; 1511 uint32_t pkt_size = m->pkt_len + HN_RNDIS_PKT_LEN; 1512 struct rndis_packet_msg *pkt; 1513 struct hn_txdesc *txd; 1514 1515 txd = hn_txd_get(txq); 1516 if (txd == NULL) 1517 break; 1518 1519 /* For small packets aggregate them in chimney buffer */ 1520 if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) { 1521 /* If this packet will not fit, then flush */ 1522 if (txq->agg_pktleft == 0 || 1523 RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) { 1524 if (hn_flush_txagg(txq, &need_sig)) 1525 goto fail; 1526 } 1527 1528 1529 pkt = hn_try_txagg(hv, txq, txd, pkt_size); 1530 if (unlikely(!pkt)) 1531 break; 1532 1533 hn_encap(pkt, queue_id, m); 1534 hn_append_to_chim(txq, pkt, m); 1535 1536 rte_pktmbuf_free(m); 1537 1538 /* if buffer is full, flush */ 1539 if (txq->agg_pktleft == 0 && 1540 hn_flush_txagg(txq, &need_sig)) 1541 goto fail; 1542 } else { 1543 /* Send any outstanding packets in buffer */ 1544 if (txq->agg_txd && hn_flush_txagg(txq, &need_sig)) 1545 goto fail; 1546 1547 pkt = txd->rndis_pkt; 1548 txd->m = m; 1549 txd->data_size = m->pkt_len; 1550 ++txd->packets; 1551 1552 hn_encap(pkt, queue_id, m); 1553 1554 ret = hn_xmit_sg(txq, txd, m, &need_sig); 1555 if (unlikely(ret != 0)) { 1556 if (ret == -EAGAIN) { 1557 PMD_TX_LOG(DEBUG, "sg channel full"); 1558 ++txq->stats.channel_full; 1559 } else { 1560 PMD_DRV_LOG(NOTICE, "sg send failed: %d", ret); 1561 ++txq->stats.errors; 1562 } 1563 hn_txd_put(txq, txd); 1564 goto fail; 1565 } 1566 } 1567 } 1568 1569 /* If partial buffer left, then try and send it. 1570 * if that fails, then reuse it on next send. 1571 */ 1572 hn_flush_txagg(txq, &need_sig); 1573 1574 fail: 1575 if (need_sig) 1576 rte_vmbus_chan_signal_tx(txq->chan); 1577 1578 return nb_tx; 1579 } 1580 1581 static uint16_t 1582 hn_recv_vf(uint16_t vf_port, const struct hn_rx_queue *rxq, 1583 struct rte_mbuf **rx_pkts, uint16_t nb_pkts) 1584 { 1585 uint16_t i, n; 1586 1587 if (unlikely(nb_pkts == 0)) 1588 return 0; 1589 1590 n = rte_eth_rx_burst(vf_port, rxq->queue_id, rx_pkts, nb_pkts); 1591 1592 /* relabel the received mbufs */ 1593 for (i = 0; i < n; i++) 1594 rx_pkts[i]->port = rxq->port_id; 1595 1596 return n; 1597 } 1598 1599 uint16_t 1600 hn_recv_pkts(void *prxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) 1601 { 1602 struct hn_rx_queue *rxq = prxq; 1603 struct hn_data *hv = rxq->hv; 1604 struct rte_eth_dev *vf_dev; 1605 uint16_t nb_rcv; 1606 1607 if (unlikely(hv->closed)) 1608 return 0; 1609 1610 /* Check for new completions (and hotplug) */ 1611 if (likely(rte_ring_count(rxq->rx_ring) < nb_pkts)) 1612 hn_process_events(hv, rxq->queue_id, 0); 1613 1614 /* Always check the vmbus path for multicast and new flows */ 1615 nb_rcv = rte_ring_sc_dequeue_burst(rxq->rx_ring, 1616 (void **)rx_pkts, nb_pkts, NULL); 1617 1618 /* If VF is available, check that as well */ 1619 rte_rwlock_read_lock(&hv->vf_lock); 1620 vf_dev = hn_get_vf_dev(hv); 1621 if (vf_dev && vf_dev->data->dev_started) 1622 nb_rcv += hn_recv_vf(vf_dev->data->port_id, rxq, 1623 rx_pkts + nb_rcv, nb_pkts - nb_rcv); 1624 1625 rte_rwlock_read_unlock(&hv->vf_lock); 1626 return nb_rcv; 1627 } 1628 1629 void 1630 hn_dev_free_queues(struct rte_eth_dev *dev) 1631 { 1632 unsigned int i; 1633 1634 for (i = 0; i < dev->data->nb_rx_queues; i++) { 1635 struct hn_rx_queue *rxq = dev->data->rx_queues[i]; 1636 1637 hn_rx_queue_free(rxq, false); 1638 dev->data->rx_queues[i] = NULL; 1639 } 1640 dev->data->nb_rx_queues = 0; 1641 1642 for (i = 0; i < dev->data->nb_tx_queues; i++) { 1643 hn_dev_tx_queue_release(dev->data->tx_queues[i]); 1644 dev->data->tx_queues[i] = NULL; 1645 } 1646 dev->data->nb_tx_queues = 0; 1647 } 1648