sfxge_tx.c revision 278221
1/*- 2 * Copyright (c) 2010-2011 Solarflare Communications, Inc. 3 * All rights reserved. 4 * 5 * This software was developed in part by Philip Paeps under contract for 6 * Solarflare Communications, Inc. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30/* Theory of operation: 31 * 32 * Tx queues allocation and mapping 33 * 34 * One Tx queue with enabled checksum offload is allocated per Rx channel 35 * (event queue). Also 2 Tx queues (one without checksum offload and one 36 * with IP checksum offload only) are allocated and bound to event queue 0. 37 * sfxge_txq_type is used as Tx queue label. 38 * 39 * So, event queue plus label mapping to Tx queue index is: 40 * if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES) 41 * else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1 42 * See sfxge_get_txq_by_label() sfxge_ev.c 43 */ 44 45#include <sys/cdefs.h> 46__FBSDID("$FreeBSD: head/sys/dev/sfxge/sfxge_tx.c 278221 2015-02-04 20:03:57Z arybchik $"); 47 48#include <sys/types.h> 49#include <sys/mbuf.h> 50#include <sys/smp.h> 51#include <sys/socket.h> 52#include <sys/sysctl.h> 53#include <sys/syslog.h> 54 55#include <net/bpf.h> 56#include <net/ethernet.h> 57#include <net/if.h> 58#include <net/if_vlan_var.h> 59 60#include <netinet/in.h> 61#include <netinet/ip.h> 62#include <netinet/ip6.h> 63#include <netinet/tcp.h> 64 65#include "common/efx.h" 66 67#include "sfxge.h" 68#include "sfxge_tx.h" 69 70/* Set the block level to ensure there is space to generate a 71 * large number of descriptors for TSO. With minimum MSS and 72 * maximum mbuf length we might need more than a ring-ful of 73 * descriptors, but this should not happen in practice except 74 * due to deliberate attack. In that case we will truncate 75 * the output at a packet boundary. Allow for a reasonable 76 * minimum MSS of 512. 77 */ 78#define SFXGE_TSO_MAX_DESC ((65535 / 512) * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1) 79#define SFXGE_TXQ_BLOCK_LEVEL(_entries) ((_entries) - SFXGE_TSO_MAX_DESC) 80 81#ifdef SFXGE_HAVE_MQ 82 83#define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max) 84static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT; 85TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max); 86SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN, 87 &sfxge_tx_dpl_get_max, 0, 88 "Maximum number of any packets in deferred packet get-list"); 89 90#define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \ 91 SFXGE_PARAM(tx_dpl_get_non_tcp_max) 92static int sfxge_tx_dpl_get_non_tcp_max = 93 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT; 94TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max); 95SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN, 96 &sfxge_tx_dpl_get_non_tcp_max, 0, 97 "Maximum number of non-TCP packets in deferred packet get-list"); 98 99#define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max) 100static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT; 101TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max); 102SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN, 103 &sfxge_tx_dpl_put_max, 0, 104 "Maximum number of any packets in deferred packet put-list"); 105 106#endif 107 108 109/* Forward declarations. */ 110static inline void sfxge_tx_qdpl_service(struct sfxge_txq *txq); 111static void sfxge_tx_qlist_post(struct sfxge_txq *txq); 112static void sfxge_tx_qunblock(struct sfxge_txq *txq); 113static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 114 const bus_dma_segment_t *dma_seg, int n_dma_seg); 115 116void 117sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) 118{ 119 unsigned int completed; 120 121 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 122 123 completed = txq->completed; 124 while (completed != txq->pending) { 125 struct sfxge_tx_mapping *stmp; 126 unsigned int id; 127 128 id = completed++ & txq->ptr_mask; 129 130 stmp = &txq->stmp[id]; 131 if (stmp->flags & TX_BUF_UNMAP) { 132 bus_dmamap_unload(txq->packet_dma_tag, stmp->map); 133 if (stmp->flags & TX_BUF_MBUF) { 134 struct mbuf *m = stmp->u.mbuf; 135 do 136 m = m_free(m); 137 while (m != NULL); 138 } else { 139 free(stmp->u.heap_buf, M_SFXGE); 140 } 141 stmp->flags = 0; 142 } 143 } 144 txq->completed = completed; 145 146 /* Check whether we need to unblock the queue. */ 147 mb(); 148 if (txq->blocked) { 149 unsigned int level; 150 151 level = txq->added - txq->completed; 152 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) 153 sfxge_tx_qunblock(txq); 154 } 155} 156 157#ifdef SFXGE_HAVE_MQ 158 159static inline unsigned int 160sfxge_is_mbuf_non_tcp(struct mbuf *mbuf) 161{ 162 /* Absense of TCP checksum flags does not mean that it is non-TCP 163 * but it should be true if user wants to achieve high throughput. 164 */ 165 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP))); 166} 167 168/* 169 * Reorder the put list and append it to the get list. 170 */ 171static void 172sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq) 173{ 174 struct sfxge_tx_dpl *stdp; 175 struct mbuf *mbuf, *get_next, **get_tailp; 176 volatile uintptr_t *putp; 177 uintptr_t put; 178 unsigned int count; 179 unsigned int non_tcp_count; 180 181 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 182 183 stdp = &txq->dpl; 184 185 /* Acquire the put list. */ 186 putp = &stdp->std_put; 187 put = atomic_readandclear_ptr(putp); 188 mbuf = (void *)put; 189 190 if (mbuf == NULL) 191 return; 192 193 /* Reverse the put list. */ 194 get_tailp = &mbuf->m_nextpkt; 195 get_next = NULL; 196 197 count = 0; 198 non_tcp_count = 0; 199 do { 200 struct mbuf *put_next; 201 202 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf); 203 put_next = mbuf->m_nextpkt; 204 mbuf->m_nextpkt = get_next; 205 get_next = mbuf; 206 mbuf = put_next; 207 208 count++; 209 } while (mbuf != NULL); 210 211 /* Append the reversed put list to the get list. */ 212 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL")); 213 *stdp->std_getp = get_next; 214 stdp->std_getp = get_tailp; 215 stdp->std_get_count += count; 216 stdp->std_get_non_tcp_count += non_tcp_count; 217} 218 219#endif /* SFXGE_HAVE_MQ */ 220 221static void 222sfxge_tx_qreap(struct sfxge_txq *txq) 223{ 224 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 225 226 txq->reaped = txq->completed; 227} 228 229static void 230sfxge_tx_qlist_post(struct sfxge_txq *txq) 231{ 232 unsigned int old_added; 233 unsigned int level; 234 int rc; 235 236 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 237 238 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0")); 239 KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC, 240 ("txq->n_pend_desc too large")); 241 KASSERT(!txq->blocked, ("txq->blocked")); 242 243 old_added = txq->added; 244 245 /* Post the fragment list. */ 246 rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc, 247 txq->reaped, &txq->added); 248 KASSERT(rc == 0, ("efx_tx_qpost() failed")); 249 250 /* If efx_tx_qpost() had to refragment, our information about 251 * buffers to free may be associated with the wrong 252 * descriptors. 253 */ 254 KASSERT(txq->added - old_added == txq->n_pend_desc, 255 ("efx_tx_qpost() refragmented descriptors")); 256 257 level = txq->added - txq->reaped; 258 KASSERT(level <= txq->entries, ("overfilled TX queue")); 259 260 /* Clear the fragment list. */ 261 txq->n_pend_desc = 0; 262 263 /* Have we reached the block level? */ 264 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) 265 return; 266 267 /* Reap, and check again */ 268 sfxge_tx_qreap(txq); 269 level = txq->added - txq->reaped; 270 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) 271 return; 272 273 txq->blocked = 1; 274 275 /* 276 * Avoid a race with completion interrupt handling that could leave 277 * the queue blocked. 278 */ 279 mb(); 280 sfxge_tx_qreap(txq); 281 level = txq->added - txq->reaped; 282 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) { 283 mb(); 284 txq->blocked = 0; 285 } 286} 287 288static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf) 289{ 290 bus_dmamap_t *used_map; 291 bus_dmamap_t map; 292 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG]; 293 unsigned int id; 294 struct sfxge_tx_mapping *stmp; 295 efx_buffer_t *desc; 296 int n_dma_seg; 297 int rc; 298 int i; 299 300 KASSERT(!txq->blocked, ("txq->blocked")); 301 302 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) 303 prefetch_read_many(mbuf->m_data); 304 305 if (txq->init_state != SFXGE_TXQ_STARTED) { 306 rc = EINTR; 307 goto reject; 308 } 309 310 /* Load the packet for DMA. */ 311 id = txq->added & txq->ptr_mask; 312 stmp = &txq->stmp[id]; 313 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, 314 mbuf, dma_seg, &n_dma_seg, 0); 315 if (rc == EFBIG) { 316 /* Try again. */ 317 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, 318 SFXGE_TX_MAPPING_MAX_SEG); 319 if (new_mbuf == NULL) 320 goto reject; 321 ++txq->collapses; 322 mbuf = new_mbuf; 323 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, 324 stmp->map, mbuf, 325 dma_seg, &n_dma_seg, 0); 326 } 327 if (rc != 0) 328 goto reject; 329 330 /* Make the packet visible to the hardware. */ 331 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); 332 333 used_map = &stmp->map; 334 335 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 336 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg); 337 if (rc < 0) 338 goto reject_mapped; 339 stmp = &txq->stmp[rc]; 340 } else { 341 /* Add the mapping to the fragment list, and set flags 342 * for the buffer. 343 */ 344 i = 0; 345 for (;;) { 346 desc = &txq->pend_desc[i]; 347 desc->eb_addr = dma_seg[i].ds_addr; 348 desc->eb_size = dma_seg[i].ds_len; 349 if (i == n_dma_seg - 1) { 350 desc->eb_eop = 1; 351 break; 352 } 353 desc->eb_eop = 0; 354 i++; 355 356 stmp->flags = 0; 357 if (__predict_false(stmp == 358 &txq->stmp[txq->ptr_mask])) 359 stmp = &txq->stmp[0]; 360 else 361 stmp++; 362 } 363 txq->n_pend_desc = n_dma_seg; 364 } 365 366 /* 367 * If the mapping required more than one descriptor 368 * then we need to associate the DMA map with the last 369 * descriptor, not the first. 370 */ 371 if (used_map != &stmp->map) { 372 map = stmp->map; 373 stmp->map = *used_map; 374 *used_map = map; 375 } 376 377 stmp->u.mbuf = mbuf; 378 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; 379 380 /* Post the fragment list. */ 381 sfxge_tx_qlist_post(txq); 382 383 return (0); 384 385reject_mapped: 386 bus_dmamap_unload(txq->packet_dma_tag, *used_map); 387reject: 388 /* Drop the packet on the floor. */ 389 m_freem(mbuf); 390 ++txq->drops; 391 392 return (rc); 393} 394 395#ifdef SFXGE_HAVE_MQ 396 397/* 398 * Drain the deferred packet list into the transmit queue. 399 */ 400static void 401sfxge_tx_qdpl_drain(struct sfxge_txq *txq) 402{ 403 struct sfxge_softc *sc; 404 struct sfxge_tx_dpl *stdp; 405 struct mbuf *mbuf, *next; 406 unsigned int count; 407 unsigned int non_tcp_count; 408 unsigned int pushed; 409 int rc; 410 411 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 412 413 sc = txq->sc; 414 stdp = &txq->dpl; 415 pushed = txq->added; 416 417 prefetch_read_many(sc->enp); 418 prefetch_read_many(txq->common); 419 420 mbuf = stdp->std_get; 421 count = stdp->std_get_count; 422 non_tcp_count = stdp->std_get_non_tcp_count; 423 424 if (count > stdp->std_get_hiwat) 425 stdp->std_get_hiwat = count; 426 427 while (count != 0) { 428 KASSERT(mbuf != NULL, ("mbuf == NULL")); 429 430 next = mbuf->m_nextpkt; 431 mbuf->m_nextpkt = NULL; 432 433 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ 434 435 if (next != NULL) 436 prefetch_read_many(next); 437 438 rc = sfxge_tx_queue_mbuf(txq, mbuf); 439 --count; 440 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); 441 mbuf = next; 442 if (rc != 0) 443 continue; 444 445 if (txq->blocked) 446 break; 447 448 /* Push the fragments to the hardware in batches. */ 449 if (txq->added - pushed >= SFXGE_TX_BATCH) { 450 efx_tx_qpush(txq->common, txq->added); 451 pushed = txq->added; 452 } 453 } 454 455 if (count == 0) { 456 KASSERT(mbuf == NULL, ("mbuf != NULL")); 457 KASSERT(non_tcp_count == 0, 458 ("inconsistent TCP/non-TCP detection")); 459 stdp->std_get = NULL; 460 stdp->std_get_count = 0; 461 stdp->std_get_non_tcp_count = 0; 462 stdp->std_getp = &stdp->std_get; 463 } else { 464 stdp->std_get = mbuf; 465 stdp->std_get_count = count; 466 stdp->std_get_non_tcp_count = non_tcp_count; 467 } 468 469 if (txq->added != pushed) 470 efx_tx_qpush(txq->common, txq->added); 471 472 KASSERT(txq->blocked || stdp->std_get_count == 0, 473 ("queue unblocked but count is non-zero")); 474} 475 476#define SFXGE_TX_QDPL_PENDING(_txq) \ 477 ((_txq)->dpl.std_put != 0) 478 479/* 480 * Service the deferred packet list. 481 * 482 * NOTE: drops the txq mutex! 483 */ 484static inline void 485sfxge_tx_qdpl_service(struct sfxge_txq *txq) 486{ 487 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 488 489 do { 490 if (SFXGE_TX_QDPL_PENDING(txq)) 491 sfxge_tx_qdpl_swizzle(txq); 492 493 if (!txq->blocked) 494 sfxge_tx_qdpl_drain(txq); 495 496 SFXGE_TXQ_UNLOCK(txq); 497 } while (SFXGE_TX_QDPL_PENDING(txq) && 498 SFXGE_TXQ_TRYLOCK(txq)); 499} 500 501/* 502 * Put a packet on the deferred packet list. 503 * 504 * If we are called with the txq lock held, we put the packet on the "get 505 * list", otherwise we atomically push it on the "put list". The swizzle 506 * function takes care of ordering. 507 * 508 * The length of the put list is bounded by SFXGE_TX_MAX_DEFFERED. We 509 * overload the csum_data field in the mbuf to keep track of this length 510 * because there is no cheap alternative to avoid races. 511 */ 512static inline int 513sfxge_tx_qdpl_put(struct sfxge_txq *txq, struct mbuf *mbuf, int locked) 514{ 515 struct sfxge_tx_dpl *stdp; 516 517 stdp = &txq->dpl; 518 519 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 520 521 if (locked) { 522 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 523 524 sfxge_tx_qdpl_swizzle(txq); 525 526 if (stdp->std_get_count >= stdp->std_get_max) { 527 txq->get_overflow++; 528 return (ENOBUFS); 529 } 530 if (sfxge_is_mbuf_non_tcp(mbuf)) { 531 if (stdp->std_get_non_tcp_count >= 532 stdp->std_get_non_tcp_max) { 533 txq->get_non_tcp_overflow++; 534 return (ENOBUFS); 535 } 536 stdp->std_get_non_tcp_count++; 537 } 538 539 *(stdp->std_getp) = mbuf; 540 stdp->std_getp = &mbuf->m_nextpkt; 541 stdp->std_get_count++; 542 } else { 543 volatile uintptr_t *putp; 544 uintptr_t old; 545 uintptr_t new; 546 unsigned old_len; 547 548 putp = &stdp->std_put; 549 new = (uintptr_t)mbuf; 550 551 do { 552 old = *putp; 553 if (old != 0) { 554 struct mbuf *mp = (struct mbuf *)old; 555 old_len = mp->m_pkthdr.csum_data; 556 } else 557 old_len = 0; 558 if (old_len >= stdp->std_put_max) { 559 atomic_add_long(&txq->put_overflow, 1); 560 return (ENOBUFS); 561 } 562 mbuf->m_pkthdr.csum_data = old_len + 1; 563 mbuf->m_nextpkt = (void *)old; 564 } while (atomic_cmpset_ptr(putp, old, new) == 0); 565 } 566 567 return (0); 568} 569 570/* 571 * Called from if_transmit - will try to grab the txq lock and enqueue to the 572 * put list if it succeeds, otherwise will push onto the defer list. 573 */ 574int 575sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) 576{ 577 int locked; 578 int rc; 579 580 if (!SFXGE_LINK_UP(txq->sc)) { 581 rc = ENETDOWN; 582 atomic_add_long(&txq->netdown_drops, 1); 583 goto fail; 584 } 585 586 /* 587 * Try to grab the txq lock. If we are able to get the lock, 588 * the packet will be appended to the "get list" of the deferred 589 * packet list. Otherwise, it will be pushed on the "put list". 590 */ 591 locked = SFXGE_TXQ_TRYLOCK(txq); 592 593 if (sfxge_tx_qdpl_put(txq, m, locked) != 0) { 594 if (locked) 595 SFXGE_TXQ_UNLOCK(txq); 596 rc = ENOBUFS; 597 goto fail; 598 } 599 600 /* 601 * Try to grab the lock again. 602 * 603 * If we are able to get the lock, we need to process the deferred 604 * packet list. If we are not able to get the lock, another thread 605 * is processing the list. 606 */ 607 if (!locked) 608 locked = SFXGE_TXQ_TRYLOCK(txq); 609 610 if (locked) { 611 /* Try to service the list. */ 612 sfxge_tx_qdpl_service(txq); 613 /* Lock has been dropped. */ 614 } 615 616 return (0); 617 618fail: 619 m_freem(m); 620 return (rc); 621} 622 623static void 624sfxge_tx_qdpl_flush(struct sfxge_txq *txq) 625{ 626 struct sfxge_tx_dpl *stdp = &txq->dpl; 627 struct mbuf *mbuf, *next; 628 629 SFXGE_TXQ_LOCK(txq); 630 631 sfxge_tx_qdpl_swizzle(txq); 632 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { 633 next = mbuf->m_nextpkt; 634 m_freem(mbuf); 635 } 636 stdp->std_get = NULL; 637 stdp->std_get_count = 0; 638 stdp->std_get_non_tcp_count = 0; 639 stdp->std_getp = &stdp->std_get; 640 641 SFXGE_TXQ_UNLOCK(txq); 642} 643 644void 645sfxge_if_qflush(struct ifnet *ifp) 646{ 647 struct sfxge_softc *sc; 648 int i; 649 650 sc = ifp->if_softc; 651 652 for (i = 0; i < SFXGE_TX_SCALE(sc); i++) 653 sfxge_tx_qdpl_flush(sc->txq[i]); 654} 655 656/* 657 * TX start -- called by the stack. 658 */ 659int 660sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) 661{ 662 struct sfxge_softc *sc; 663 struct sfxge_txq *txq; 664 int rc; 665 666 sc = (struct sfxge_softc *)ifp->if_softc; 667 668 KASSERT(ifp->if_flags & IFF_UP, ("interface not up")); 669 670 /* Pick the desired transmit queue. */ 671 if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) { 672 int index = 0; 673 674 /* check if flowid is set */ 675 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 676 uint32_t hash = m->m_pkthdr.flowid; 677 678 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX]; 679 } 680 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; 681 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { 682 txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; 683 } else { 684 txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; 685 } 686 687 rc = sfxge_tx_packet_add(txq, m); 688 689 return (rc); 690} 691 692#else /* !SFXGE_HAVE_MQ */ 693 694static void sfxge_if_start_locked(struct ifnet *ifp) 695{ 696 struct sfxge_softc *sc = ifp->if_softc; 697 struct sfxge_txq *txq; 698 struct mbuf *mbuf; 699 unsigned int pushed[SFXGE_TXQ_NTYPES]; 700 unsigned int q_index; 701 702 if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) != 703 IFF_DRV_RUNNING) 704 return; 705 706 if (!sc->port.link_up) 707 return; 708 709 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) { 710 txq = sc->txq[q_index]; 711 pushed[q_index] = txq->added; 712 } 713 714 while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) { 715 IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf); 716 if (mbuf == NULL) 717 break; 718 719 ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */ 720 721 /* Pick the desired transmit queue. */ 722 if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) 723 q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM; 724 else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP) 725 q_index = SFXGE_TXQ_IP_CKSUM; 726 else 727 q_index = SFXGE_TXQ_NON_CKSUM; 728 txq = sc->txq[q_index]; 729 730 if (sfxge_tx_queue_mbuf(txq, mbuf) != 0) 731 continue; 732 733 if (txq->blocked) { 734 ifp->if_drv_flags |= IFF_DRV_OACTIVE; 735 break; 736 } 737 738 /* Push the fragments to the hardware in batches. */ 739 if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) { 740 efx_tx_qpush(txq->common, txq->added); 741 pushed[q_index] = txq->added; 742 } 743 } 744 745 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) { 746 txq = sc->txq[q_index]; 747 if (txq->added != pushed[q_index]) 748 efx_tx_qpush(txq->common, txq->added); 749 } 750} 751 752void sfxge_if_start(struct ifnet *ifp) 753{ 754 struct sfxge_softc *sc = ifp->if_softc; 755 756 SFXGE_TXQ_LOCK(sc->txq[0]); 757 sfxge_if_start_locked(ifp); 758 SFXGE_TXQ_UNLOCK(sc->txq[0]); 759} 760 761static inline void 762sfxge_tx_qdpl_service(struct sfxge_txq *txq) 763{ 764 struct ifnet *ifp = txq->sc->ifnet; 765 766 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 767 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE; 768 sfxge_if_start_locked(ifp); 769 SFXGE_TXQ_UNLOCK(txq); 770} 771 772#endif /* SFXGE_HAVE_MQ */ 773 774/* 775 * Software "TSO". Not quite as good as doing it in hardware, but 776 * still faster than segmenting in the stack. 777 */ 778 779struct sfxge_tso_state { 780 /* Output position */ 781 unsigned out_len; /* Remaining length in current segment */ 782 unsigned seqnum; /* Current sequence number */ 783 unsigned packet_space; /* Remaining space in current packet */ 784 785 /* Input position */ 786 unsigned dma_seg_i; /* Current DMA segment number */ 787 uint64_t dma_addr; /* DMA address of current position */ 788 unsigned in_len; /* Remaining length in current mbuf */ 789 790 const struct mbuf *mbuf; /* Input mbuf (head of chain) */ 791 u_short protocol; /* Network protocol (after VLAN decap) */ 792 ssize_t nh_off; /* Offset of network header */ 793 ssize_t tcph_off; /* Offset of TCP header */ 794 unsigned header_len; /* Number of bytes of header */ 795 int full_packet_size; /* Number of bytes to put in each outgoing 796 * segment */ 797}; 798 799static inline const struct ip *tso_iph(const struct sfxge_tso_state *tso) 800{ 801 KASSERT(tso->protocol == htons(ETHERTYPE_IP), 802 ("tso_iph() in non-IPv4 state")); 803 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); 804} 805static inline const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) 806{ 807 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 808 ("tso_ip6h() in non-IPv6 state")); 809 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); 810} 811static inline const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) 812{ 813 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); 814} 815 816/* Size of preallocated TSO header buffers. Larger blocks must be 817 * allocated from the heap. 818 */ 819#define TSOH_STD_SIZE 128 820 821/* At most half the descriptors in the queue at any time will refer to 822 * a TSO header buffer, since they must always be followed by a 823 * payload descriptor referring to an mbuf. 824 */ 825#define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) 826#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 827#define TSOH_PAGE_COUNT(_txq_entries) \ 828 ((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE) 829 830static int tso_init(struct sfxge_txq *txq) 831{ 832 struct sfxge_softc *sc = txq->sc; 833 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); 834 int i, rc; 835 836 /* Allocate TSO header buffers */ 837 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), 838 M_SFXGE, M_WAITOK); 839 840 for (i = 0; i < tsoh_page_count; i++) { 841 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); 842 if (rc != 0) 843 goto fail; 844 } 845 846 return (0); 847 848fail: 849 while (i-- > 0) 850 sfxge_dma_free(&txq->tsoh_buffer[i]); 851 free(txq->tsoh_buffer, M_SFXGE); 852 txq->tsoh_buffer = NULL; 853 return (rc); 854} 855 856static void tso_fini(struct sfxge_txq *txq) 857{ 858 int i; 859 860 if (txq->tsoh_buffer != NULL) { 861 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) 862 sfxge_dma_free(&txq->tsoh_buffer[i]); 863 free(txq->tsoh_buffer, M_SFXGE); 864 } 865} 866 867static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf) 868{ 869 struct ether_header *eh = mtod(mbuf, struct ether_header *); 870 871 tso->mbuf = mbuf; 872 873 /* Find network protocol and header */ 874 tso->protocol = eh->ether_type; 875 if (tso->protocol == htons(ETHERTYPE_VLAN)) { 876 struct ether_vlan_header *veh = 877 mtod(mbuf, struct ether_vlan_header *); 878 tso->protocol = veh->evl_proto; 879 tso->nh_off = sizeof(*veh); 880 } else { 881 tso->nh_off = sizeof(*eh); 882 } 883 884 /* Find TCP header */ 885 if (tso->protocol == htons(ETHERTYPE_IP)) { 886 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, 887 ("TSO required on non-TCP packet")); 888 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; 889 } else { 890 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 891 ("TSO required on non-IP packet")); 892 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, 893 ("TSO required on non-TCP packet")); 894 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); 895 } 896 897 tso->header_len = tso->tcph_off + 4 * tso_tcph(tso)->th_off; 898 tso->full_packet_size = tso->header_len + mbuf->m_pkthdr.tso_segsz; 899 900 tso->seqnum = ntohl(tso_tcph(tso)->th_seq); 901 902 /* These flags must not be duplicated */ 903 KASSERT(!(tso_tcph(tso)->th_flags & (TH_URG | TH_SYN | TH_RST)), 904 ("incompatible TCP flag on TSO packet")); 905 906 tso->out_len = mbuf->m_pkthdr.len - tso->header_len; 907} 908 909/* 910 * tso_fill_packet_with_fragment - form descriptors for the current fragment 911 * 912 * Form descriptors for the current fragment, until we reach the end 913 * of fragment or end-of-packet. Return 0 on success, 1 if not enough 914 * space. 915 */ 916static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, 917 struct sfxge_tso_state *tso) 918{ 919 efx_buffer_t *desc; 920 int n; 921 922 if (tso->in_len == 0 || tso->packet_space == 0) 923 return; 924 925 KASSERT(tso->in_len > 0, ("TSO input length went negative")); 926 KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); 927 928 n = min(tso->in_len, tso->packet_space); 929 930 tso->packet_space -= n; 931 tso->out_len -= n; 932 tso->in_len -= n; 933 934 desc = &txq->pend_desc[txq->n_pend_desc++]; 935 desc->eb_addr = tso->dma_addr; 936 desc->eb_size = n; 937 desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0; 938 939 tso->dma_addr += n; 940} 941 942/* Callback from bus_dmamap_load() for long TSO headers. */ 943static void tso_map_long_header(void *dma_addr_ret, 944 bus_dma_segment_t *segs, int nseg, 945 int error) 946{ 947 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && 948 __predict_true(nseg == 1)) ? 949 segs->ds_addr : 0); 950} 951 952/* 953 * tso_start_new_packet - generate a new header and prepare for the new packet 954 * 955 * Generate a new header and prepare for the new packet. Return 0 on 956 * success, or an error code if failed to alloc header. 957 */ 958static int tso_start_new_packet(struct sfxge_txq *txq, 959 struct sfxge_tso_state *tso, 960 unsigned int id) 961{ 962 struct sfxge_tx_mapping *stmp = &txq->stmp[id]; 963 struct tcphdr *tsoh_th; 964 unsigned ip_length; 965 caddr_t header; 966 uint64_t dma_addr; 967 bus_dmamap_t map; 968 efx_buffer_t *desc; 969 int rc; 970 971 /* Allocate a DMA-mapped header buffer. */ 972 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { 973 unsigned int page_index = (id / 2) / TSOH_PER_PAGE; 974 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; 975 976 header = (txq->tsoh_buffer[page_index].esm_base + 977 buf_index * TSOH_STD_SIZE); 978 dma_addr = (txq->tsoh_buffer[page_index].esm_addr + 979 buf_index * TSOH_STD_SIZE); 980 map = txq->tsoh_buffer[page_index].esm_map; 981 982 stmp->flags = 0; 983 } else { 984 /* We cannot use bus_dmamem_alloc() as that may sleep */ 985 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); 986 if (__predict_false(!header)) 987 return (ENOMEM); 988 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, 989 header, tso->header_len, 990 tso_map_long_header, &dma_addr, 991 BUS_DMA_NOWAIT); 992 if (__predict_false(dma_addr == 0)) { 993 if (rc == 0) { 994 /* Succeeded but got >1 segment */ 995 bus_dmamap_unload(txq->packet_dma_tag, 996 stmp->map); 997 rc = EINVAL; 998 } 999 free(header, M_SFXGE); 1000 return (rc); 1001 } 1002 map = stmp->map; 1003 1004 txq->tso_long_headers++; 1005 stmp->u.heap_buf = header; 1006 stmp->flags = TX_BUF_UNMAP; 1007 } 1008 1009 tsoh_th = (struct tcphdr *)(header + tso->tcph_off); 1010 1011 /* Copy and update the headers. */ 1012 m_copydata(tso->mbuf, 0, tso->header_len, header); 1013 1014 tsoh_th->th_seq = htonl(tso->seqnum); 1015 tso->seqnum += tso->mbuf->m_pkthdr.tso_segsz; 1016 if (tso->out_len > tso->mbuf->m_pkthdr.tso_segsz) { 1017 /* This packet will not finish the TSO burst. */ 1018 ip_length = tso->full_packet_size - tso->nh_off; 1019 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); 1020 } else { 1021 /* This packet will be the last in the TSO burst. */ 1022 ip_length = tso->header_len - tso->nh_off + tso->out_len; 1023 } 1024 1025 if (tso->protocol == htons(ETHERTYPE_IP)) { 1026 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); 1027 tsoh_iph->ip_len = htons(ip_length); 1028 /* XXX We should increment ip_id, but FreeBSD doesn't 1029 * currently allocate extra IDs for multiple segments. 1030 */ 1031 } else { 1032 struct ip6_hdr *tsoh_iph = 1033 (struct ip6_hdr *)(header + tso->nh_off); 1034 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); 1035 } 1036 1037 /* Make the header visible to the hardware. */ 1038 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); 1039 1040 tso->packet_space = tso->mbuf->m_pkthdr.tso_segsz; 1041 txq->tso_packets++; 1042 1043 /* Form a descriptor for this header. */ 1044 desc = &txq->pend_desc[txq->n_pend_desc++]; 1045 desc->eb_addr = dma_addr; 1046 desc->eb_size = tso->header_len; 1047 desc->eb_eop = 0; 1048 1049 return (0); 1050} 1051 1052static int 1053sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 1054 const bus_dma_segment_t *dma_seg, int n_dma_seg) 1055{ 1056 struct sfxge_tso_state tso; 1057 unsigned int id, next_id; 1058 unsigned skipped = 0; 1059 1060 tso_start(&tso, mbuf); 1061 1062 while (dma_seg->ds_len + skipped <= tso.header_len) { 1063 skipped += dma_seg->ds_len; 1064 --n_dma_seg; 1065 KASSERT(n_dma_seg, ("no payload found in TSO packet")); 1066 ++dma_seg; 1067 } 1068 tso.in_len = dma_seg->ds_len + (tso.header_len - skipped); 1069 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); 1070 1071 id = txq->added & txq->ptr_mask; 1072 if (__predict_false(tso_start_new_packet(txq, &tso, id))) 1073 return (-1); 1074 1075 while (1) { 1076 id = (id + 1) & txq->ptr_mask; 1077 tso_fill_packet_with_fragment(txq, &tso); 1078 1079 /* Move onto the next fragment? */ 1080 if (tso.in_len == 0) { 1081 --n_dma_seg; 1082 if (n_dma_seg == 0) 1083 break; 1084 ++dma_seg; 1085 tso.in_len = dma_seg->ds_len; 1086 tso.dma_addr = dma_seg->ds_addr; 1087 } 1088 1089 /* End of packet? */ 1090 if (tso.packet_space == 0) { 1091 /* If the queue is now full due to tiny MSS, 1092 * or we can't create another header, discard 1093 * the remainder of the input mbuf but do not 1094 * roll back the work we have done. 1095 */ 1096 if (txq->n_pend_desc > 1097 SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG)) 1098 break; 1099 next_id = (id + 1) & txq->ptr_mask; 1100 if (__predict_false(tso_start_new_packet(txq, &tso, 1101 next_id))) 1102 break; 1103 id = next_id; 1104 } 1105 } 1106 1107 txq->tso_bursts++; 1108 return (id); 1109} 1110 1111static void 1112sfxge_tx_qunblock(struct sfxge_txq *txq) 1113{ 1114 struct sfxge_softc *sc; 1115 struct sfxge_evq *evq; 1116 1117 sc = txq->sc; 1118 evq = sc->evq[txq->evq_index]; 1119 1120 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 1121 1122 if (txq->init_state != SFXGE_TXQ_STARTED) 1123 return; 1124 1125 SFXGE_TXQ_LOCK(txq); 1126 1127 if (txq->blocked) { 1128 unsigned int level; 1129 1130 level = txq->added - txq->completed; 1131 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) 1132 txq->blocked = 0; 1133 } 1134 1135 sfxge_tx_qdpl_service(txq); 1136 /* note: lock has been dropped */ 1137} 1138 1139void 1140sfxge_tx_qflush_done(struct sfxge_txq *txq) 1141{ 1142 1143 txq->flush_state = SFXGE_FLUSH_DONE; 1144} 1145 1146static void 1147sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) 1148{ 1149 struct sfxge_txq *txq; 1150 struct sfxge_evq *evq; 1151 unsigned int count; 1152 1153 txq = sc->txq[index]; 1154 evq = sc->evq[txq->evq_index]; 1155 1156 SFXGE_TXQ_LOCK(txq); 1157 1158 KASSERT(txq->init_state == SFXGE_TXQ_STARTED, 1159 ("txq->init_state != SFXGE_TXQ_STARTED")); 1160 1161 txq->init_state = SFXGE_TXQ_INITIALIZED; 1162 txq->flush_state = SFXGE_FLUSH_PENDING; 1163 1164 /* Flush the transmit queue. */ 1165 efx_tx_qflush(txq->common); 1166 1167 SFXGE_TXQ_UNLOCK(txq); 1168 1169 count = 0; 1170 do { 1171 /* Spin for 100ms. */ 1172 DELAY(100000); 1173 1174 if (txq->flush_state != SFXGE_FLUSH_PENDING) 1175 break; 1176 } while (++count < 20); 1177 1178 SFXGE_EVQ_LOCK(evq); 1179 SFXGE_TXQ_LOCK(txq); 1180 1181 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, 1182 ("txq->flush_state == SFXGE_FLUSH_FAILED")); 1183 1184 txq->flush_state = SFXGE_FLUSH_DONE; 1185 1186 txq->blocked = 0; 1187 txq->pending = txq->added; 1188 1189 sfxge_tx_qcomplete(txq, evq); 1190 KASSERT(txq->completed == txq->added, 1191 ("txq->completed != txq->added")); 1192 1193 sfxge_tx_qreap(txq); 1194 KASSERT(txq->reaped == txq->completed, 1195 ("txq->reaped != txq->completed")); 1196 1197 txq->added = 0; 1198 txq->pending = 0; 1199 txq->completed = 0; 1200 txq->reaped = 0; 1201 1202 /* Destroy the common code transmit queue. */ 1203 efx_tx_qdestroy(txq->common); 1204 txq->common = NULL; 1205 1206 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1207 EFX_TXQ_NBUFS(sc->txq_entries)); 1208 1209 SFXGE_EVQ_UNLOCK(evq); 1210 SFXGE_TXQ_UNLOCK(txq); 1211} 1212 1213static int 1214sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) 1215{ 1216 struct sfxge_txq *txq; 1217 efsys_mem_t *esmp; 1218 uint16_t flags; 1219 struct sfxge_evq *evq; 1220 int rc; 1221 1222 txq = sc->txq[index]; 1223 esmp = &txq->mem; 1224 evq = sc->evq[txq->evq_index]; 1225 1226 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1227 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1228 KASSERT(evq->init_state == SFXGE_EVQ_STARTED, 1229 ("evq->init_state != SFXGE_EVQ_STARTED")); 1230 1231 /* Program the buffer table. */ 1232 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, 1233 EFX_TXQ_NBUFS(sc->txq_entries))) != 0) 1234 return (rc); 1235 1236 /* Determine the kind of queue we are creating. */ 1237 switch (txq->type) { 1238 case SFXGE_TXQ_NON_CKSUM: 1239 flags = 0; 1240 break; 1241 case SFXGE_TXQ_IP_CKSUM: 1242 flags = EFX_CKSUM_IPV4; 1243 break; 1244 case SFXGE_TXQ_IP_TCP_UDP_CKSUM: 1245 flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP; 1246 break; 1247 default: 1248 KASSERT(0, ("Impossible TX queue")); 1249 flags = 0; 1250 break; 1251 } 1252 1253 /* Create the common code transmit queue. */ 1254 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1255 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1256 &txq->common)) != 0) 1257 goto fail; 1258 1259 SFXGE_TXQ_LOCK(txq); 1260 1261 /* Enable the transmit queue. */ 1262 efx_tx_qenable(txq->common); 1263 1264 txq->init_state = SFXGE_TXQ_STARTED; 1265 1266 SFXGE_TXQ_UNLOCK(txq); 1267 1268 return (0); 1269 1270fail: 1271 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1272 EFX_TXQ_NBUFS(sc->txq_entries)); 1273 return (rc); 1274} 1275 1276void 1277sfxge_tx_stop(struct sfxge_softc *sc) 1278{ 1279 const efx_nic_cfg_t *encp; 1280 int index; 1281 1282 index = SFXGE_TX_SCALE(sc); 1283 while (--index >= 0) 1284 sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1285 1286 sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM); 1287 1288 encp = efx_nic_cfg_get(sc->enp); 1289 sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM); 1290 1291 /* Tear down the transmit module */ 1292 efx_tx_fini(sc->enp); 1293} 1294 1295int 1296sfxge_tx_start(struct sfxge_softc *sc) 1297{ 1298 int index; 1299 int rc; 1300 1301 /* Initialize the common code transmit module. */ 1302 if ((rc = efx_tx_init(sc->enp)) != 0) 1303 return (rc); 1304 1305 if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_NON_CKSUM)) != 0) 1306 goto fail; 1307 1308 if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_CKSUM)) != 0) 1309 goto fail2; 1310 1311 for (index = 0; index < SFXGE_TX_SCALE(sc); index++) { 1312 if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + 1313 index)) != 0) 1314 goto fail3; 1315 } 1316 1317 return (0); 1318 1319fail3: 1320 while (--index >= 0) 1321 sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1322 1323 sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM); 1324 1325fail2: 1326 sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM); 1327 1328fail: 1329 efx_tx_fini(sc->enp); 1330 1331 return (rc); 1332} 1333 1334/** 1335 * Destroy a transmit queue. 1336 */ 1337static void 1338sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) 1339{ 1340 struct sfxge_txq *txq; 1341 unsigned int nmaps; 1342 1343 txq = sc->txq[index]; 1344 1345 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1346 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1347 1348 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) 1349 tso_fini(txq); 1350 1351 /* Free the context arrays. */ 1352 free(txq->pend_desc, M_SFXGE); 1353 nmaps = sc->txq_entries; 1354 while (nmaps-- != 0) 1355 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1356 free(txq->stmp, M_SFXGE); 1357 1358 /* Release DMA memory mapping. */ 1359 sfxge_dma_free(&txq->mem); 1360 1361 sc->txq[index] = NULL; 1362 1363#ifdef SFXGE_HAVE_MQ 1364 SFXGE_TXQ_LOCK_DESTROY(txq); 1365#endif 1366 1367 free(txq, M_SFXGE); 1368} 1369 1370static int 1371sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, 1372 enum sfxge_txq_type type, unsigned int evq_index) 1373{ 1374 char name[16]; 1375 struct sysctl_oid *txq_node; 1376 struct sfxge_txq *txq; 1377 struct sfxge_evq *evq; 1378#ifdef SFXGE_HAVE_MQ 1379 struct sfxge_tx_dpl *stdp; 1380#endif 1381 efsys_mem_t *esmp; 1382 unsigned int nmaps; 1383 int rc; 1384 1385 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); 1386 txq->sc = sc; 1387 txq->entries = sc->txq_entries; 1388 txq->ptr_mask = txq->entries - 1; 1389 1390 sc->txq[txq_index] = txq; 1391 esmp = &txq->mem; 1392 1393 evq = sc->evq[evq_index]; 1394 1395 /* Allocate and zero DMA space for the descriptor ring. */ 1396 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) 1397 return (rc); 1398 (void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(sc->txq_entries)); 1399 1400 /* Allocate buffer table entries. */ 1401 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), 1402 &txq->buf_base_id); 1403 1404 /* Create a DMA tag for packet mappings. */ 1405 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000, 1406 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, 1407 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL, 1408 &txq->packet_dma_tag) != 0) { 1409 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); 1410 rc = ENOMEM; 1411 goto fail; 1412 } 1413 1414 /* Allocate pending descriptor array for batching writes. */ 1415 txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries, 1416 M_SFXGE, M_ZERO | M_WAITOK); 1417 1418 /* Allocate and initialise mbuf DMA mapping array. */ 1419 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, 1420 M_SFXGE, M_ZERO | M_WAITOK); 1421 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { 1422 rc = bus_dmamap_create(txq->packet_dma_tag, 0, 1423 &txq->stmp[nmaps].map); 1424 if (rc != 0) 1425 goto fail2; 1426 } 1427 1428 snprintf(name, sizeof(name), "%u", txq_index); 1429 txq_node = SYSCTL_ADD_NODE( 1430 device_get_sysctl_ctx(sc->dev), 1431 SYSCTL_CHILDREN(sc->txqs_node), 1432 OID_AUTO, name, CTLFLAG_RD, NULL, ""); 1433 if (txq_node == NULL) { 1434 rc = ENOMEM; 1435 goto fail_txq_node; 1436 } 1437 1438 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && 1439 (rc = tso_init(txq)) != 0) 1440 goto fail3; 1441 1442#ifdef SFXGE_HAVE_MQ 1443 if (sfxge_tx_dpl_get_max <= 0) { 1444 log(LOG_ERR, "%s=%d must be greater than 0", 1445 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); 1446 rc = EINVAL; 1447 goto fail_tx_dpl_get_max; 1448 } 1449 if (sfxge_tx_dpl_get_non_tcp_max <= 0) { 1450 log(LOG_ERR, "%s=%d must be greater than 0", 1451 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, 1452 sfxge_tx_dpl_get_non_tcp_max); 1453 rc = EINVAL; 1454 goto fail_tx_dpl_get_max; 1455 } 1456 if (sfxge_tx_dpl_put_max < 0) { 1457 log(LOG_ERR, "%s=%d must be greater or equal to 0", 1458 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); 1459 rc = EINVAL; 1460 goto fail_tx_dpl_put_max; 1461 } 1462 1463 /* Initialize the deferred packet list. */ 1464 stdp = &txq->dpl; 1465 stdp->std_put_max = sfxge_tx_dpl_put_max; 1466 stdp->std_get_max = sfxge_tx_dpl_get_max; 1467 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; 1468 stdp->std_getp = &stdp->std_get; 1469 1470 SFXGE_TXQ_LOCK_INIT(txq, "txq"); 1471 1472 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev), 1473 SYSCTL_CHILDREN(txq_node), OID_AUTO, 1474 "dpl_get_count", CTLFLAG_RD | CTLFLAG_STATS, 1475 &stdp->std_get_count, 0, ""); 1476 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev), 1477 SYSCTL_CHILDREN(txq_node), OID_AUTO, 1478 "dpl_get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, 1479 &stdp->std_get_non_tcp_count, 0, ""); 1480 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev), 1481 SYSCTL_CHILDREN(txq_node), OID_AUTO, 1482 "dpl_get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1483 &stdp->std_get_hiwat, 0, ""); 1484#endif 1485 1486 txq->type = type; 1487 txq->evq_index = evq_index; 1488 txq->txq_index = txq_index; 1489 txq->init_state = SFXGE_TXQ_INITIALIZED; 1490 1491 return (0); 1492 1493fail_tx_dpl_put_max: 1494fail_tx_dpl_get_max: 1495fail3: 1496fail_txq_node: 1497 free(txq->pend_desc, M_SFXGE); 1498fail2: 1499 while (nmaps-- != 0) 1500 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1501 free(txq->stmp, M_SFXGE); 1502 bus_dma_tag_destroy(txq->packet_dma_tag); 1503 1504fail: 1505 sfxge_dma_free(esmp); 1506 1507 return (rc); 1508} 1509 1510static const struct { 1511 const char *name; 1512 size_t offset; 1513} sfxge_tx_stats[] = { 1514#define SFXGE_TX_STAT(name, member) \ 1515 { #name, offsetof(struct sfxge_txq, member) } 1516 SFXGE_TX_STAT(tso_bursts, tso_bursts), 1517 SFXGE_TX_STAT(tso_packets, tso_packets), 1518 SFXGE_TX_STAT(tso_long_headers, tso_long_headers), 1519 SFXGE_TX_STAT(tx_collapses, collapses), 1520 SFXGE_TX_STAT(tx_drops, drops), 1521 SFXGE_TX_STAT(tx_get_overflow, get_overflow), 1522 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow), 1523 SFXGE_TX_STAT(tx_put_overflow, put_overflow), 1524 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops), 1525}; 1526 1527static int 1528sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS) 1529{ 1530 struct sfxge_softc *sc = arg1; 1531 unsigned int id = arg2; 1532 unsigned long sum; 1533 unsigned int index; 1534 1535 /* Sum across all TX queues */ 1536 sum = 0; 1537 for (index = 0; 1538 index < SFXGE_TXQ_IP_TCP_UDP_CKSUM + SFXGE_TX_SCALE(sc); 1539 index++) 1540 sum += *(unsigned long *)((caddr_t)sc->txq[index] + 1541 sfxge_tx_stats[id].offset); 1542 1543 return (SYSCTL_OUT(req, &sum, sizeof(sum))); 1544} 1545 1546static void 1547sfxge_tx_stat_init(struct sfxge_softc *sc) 1548{ 1549 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1550 struct sysctl_oid_list *stat_list; 1551 unsigned int id; 1552 1553 stat_list = SYSCTL_CHILDREN(sc->stats_node); 1554 1555 for (id = 0; 1556 id < sizeof(sfxge_tx_stats) / sizeof(sfxge_tx_stats[0]); 1557 id++) { 1558 SYSCTL_ADD_PROC( 1559 ctx, stat_list, 1560 OID_AUTO, sfxge_tx_stats[id].name, 1561 CTLTYPE_ULONG|CTLFLAG_RD, 1562 sc, id, sfxge_tx_stat_handler, "LU", 1563 ""); 1564 } 1565} 1566 1567void 1568sfxge_tx_fini(struct sfxge_softc *sc) 1569{ 1570 int index; 1571 1572 index = SFXGE_TX_SCALE(sc); 1573 while (--index >= 0) 1574 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1575 1576 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 1577 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 1578} 1579 1580 1581int 1582sfxge_tx_init(struct sfxge_softc *sc) 1583{ 1584 struct sfxge_intr *intr; 1585 int index; 1586 int rc; 1587 1588 intr = &sc->intr; 1589 1590 KASSERT(intr->state == SFXGE_INTR_INITIALIZED, 1591 ("intr->state != SFXGE_INTR_INITIALIZED")); 1592 1593 sc->txqs_node = SYSCTL_ADD_NODE( 1594 device_get_sysctl_ctx(sc->dev), 1595 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), 1596 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); 1597 if (sc->txqs_node == NULL) { 1598 rc = ENOMEM; 1599 goto fail_txq_node; 1600 } 1601 1602 /* Initialize the transmit queues */ 1603 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, 1604 SFXGE_TXQ_NON_CKSUM, 0)) != 0) 1605 goto fail; 1606 1607 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, 1608 SFXGE_TXQ_IP_CKSUM, 0)) != 0) 1609 goto fail2; 1610 1611 for (index = 0; index < SFXGE_TX_SCALE(sc); index++) { 1612 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index, 1613 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) 1614 goto fail3; 1615 } 1616 1617 sfxge_tx_stat_init(sc); 1618 1619 return (0); 1620 1621fail3: 1622 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 1623 1624 while (--index >= 0) 1625 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1626 1627fail2: 1628 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 1629 1630fail: 1631fail_txq_node: 1632 return (rc); 1633} 1634