sfxge_tx.c revision 342451
1/*- 2 * Copyright (c) 2010-2016 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 are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright notice, 12 * this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright notice, 14 * this list of conditions and the following disclaimer in the documentation 15 * and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 18 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, 19 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR 21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR 26 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, 27 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 * 29 * The views and conclusions contained in the software and documentation are 30 * those of the authors and should not be interpreted as representing official 31 * policies, either expressed or implied, of the FreeBSD Project. 32 */ 33 34/* Theory of operation: 35 * 36 * Tx queues allocation and mapping 37 * 38 * One Tx queue with enabled checksum offload is allocated per Rx channel 39 * (event queue). Also 2 Tx queues (one without checksum offload and one 40 * with IP checksum offload only) are allocated and bound to event queue 0. 41 * sfxge_txq_type is used as Tx queue label. 42 * 43 * So, event queue plus label mapping to Tx queue index is: 44 * if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES) 45 * else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1 46 * See sfxge_get_txq_by_label() sfxge_ev.c 47 */ 48 49#include <sys/cdefs.h> 50__FBSDID("$FreeBSD: stable/11/sys/dev/sfxge/sfxge_tx.c 342451 2018-12-25 07:33:45Z arybchik $"); 51 52#include "opt_rss.h" 53 54#include <sys/param.h> 55#include <sys/malloc.h> 56#include <sys/mbuf.h> 57#include <sys/smp.h> 58#include <sys/socket.h> 59#include <sys/sysctl.h> 60#include <sys/syslog.h> 61#include <sys/limits.h> 62 63#include <net/bpf.h> 64#include <net/ethernet.h> 65#include <net/if.h> 66#include <net/if_vlan_var.h> 67 68#include <netinet/in.h> 69#include <netinet/ip.h> 70#include <netinet/ip6.h> 71#include <netinet/tcp.h> 72 73#ifdef RSS 74#include <net/rss_config.h> 75#endif 76 77#include "common/efx.h" 78 79#include "sfxge.h" 80#include "sfxge_tx.h" 81 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#define SFXGE_PARAM_TSO_FW_ASSISTED SFXGE_PARAM(tso_fw_assisted) 107static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2); 108TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted); 109SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN, 110 &sfxge_tso_fw_assisted, 0, 111 "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware"); 112 113 114static const struct { 115 const char *name; 116 size_t offset; 117} sfxge_tx_stats[] = { 118#define SFXGE_TX_STAT(name, member) \ 119 { #name, offsetof(struct sfxge_txq, member) } 120 SFXGE_TX_STAT(tso_bursts, tso_bursts), 121 SFXGE_TX_STAT(tso_packets, tso_packets), 122 SFXGE_TX_STAT(tso_long_headers, tso_long_headers), 123 SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many), 124 SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc), 125 SFXGE_TX_STAT(tx_collapses, collapses), 126 SFXGE_TX_STAT(tx_drops, drops), 127 SFXGE_TX_STAT(tx_get_overflow, get_overflow), 128 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow), 129 SFXGE_TX_STAT(tx_put_overflow, put_overflow), 130 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops), 131}; 132 133 134/* Forward declarations. */ 135static void sfxge_tx_qdpl_service(struct sfxge_txq *txq); 136static void sfxge_tx_qlist_post(struct sfxge_txq *txq); 137static void sfxge_tx_qunblock(struct sfxge_txq *txq); 138static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 139 const bus_dma_segment_t *dma_seg, int n_dma_seg, 140 int vlan_tagged); 141 142static int 143sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf) 144{ 145 uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ? 146 mbuf->m_pkthdr.ether_vtag : 147 0); 148 149 if (this_tag == txq->hw_vlan_tci) 150 return (0); 151 152 efx_tx_qdesc_vlantci_create(txq->common, 153 bswap16(this_tag), 154 &txq->pend_desc[0]); 155 txq->n_pend_desc = 1; 156 txq->hw_vlan_tci = this_tag; 157 return (1); 158} 159 160static inline void 161sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp) 162{ 163 KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0")); 164 if (__predict_false(*pstmp == 165 &txq->stmp[txq->ptr_mask])) 166 *pstmp = &txq->stmp[0]; 167 else 168 (*pstmp)++; 169} 170 171 172void 173sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) 174{ 175 unsigned int completed; 176 177 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 178 179 completed = txq->completed; 180 while (completed != txq->pending) { 181 struct sfxge_tx_mapping *stmp; 182 unsigned int id; 183 184 id = completed++ & txq->ptr_mask; 185 186 stmp = &txq->stmp[id]; 187 if (stmp->flags & TX_BUF_UNMAP) { 188 bus_dmamap_unload(txq->packet_dma_tag, stmp->map); 189 if (stmp->flags & TX_BUF_MBUF) { 190 struct mbuf *m = stmp->u.mbuf; 191 do 192 m = m_free(m); 193 while (m != NULL); 194 } else { 195 free(stmp->u.heap_buf, M_SFXGE); 196 } 197 stmp->flags = 0; 198 } 199 } 200 txq->completed = completed; 201 202 /* Check whether we need to unblock the queue. */ 203 mb(); 204 if (txq->blocked) { 205 unsigned int level; 206 207 level = txq->added - txq->completed; 208 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) 209 sfxge_tx_qunblock(txq); 210 } 211} 212 213static unsigned int 214sfxge_is_mbuf_non_tcp(struct mbuf *mbuf) 215{ 216 /* Absence of TCP checksum flags does not mean that it is non-TCP 217 * but it should be true if user wants to achieve high throughput. 218 */ 219 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP))); 220} 221 222/* 223 * Reorder the put list and append it to the get list. 224 */ 225static void 226sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq) 227{ 228 struct sfxge_tx_dpl *stdp; 229 struct mbuf *mbuf, *get_next, **get_tailp; 230 volatile uintptr_t *putp; 231 uintptr_t put; 232 unsigned int count; 233 unsigned int non_tcp_count; 234 235 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 236 237 stdp = &txq->dpl; 238 239 /* Acquire the put list. */ 240 putp = &stdp->std_put; 241 put = atomic_readandclear_ptr(putp); 242 mbuf = (void *)put; 243 244 if (mbuf == NULL) 245 return; 246 247 /* Reverse the put list. */ 248 get_tailp = &mbuf->m_nextpkt; 249 get_next = NULL; 250 251 count = 0; 252 non_tcp_count = 0; 253 do { 254 struct mbuf *put_next; 255 256 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf); 257 put_next = mbuf->m_nextpkt; 258 mbuf->m_nextpkt = get_next; 259 get_next = mbuf; 260 mbuf = put_next; 261 262 count++; 263 } while (mbuf != NULL); 264 265 if (count > stdp->std_put_hiwat) 266 stdp->std_put_hiwat = count; 267 268 /* Append the reversed put list to the get list. */ 269 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL")); 270 *stdp->std_getp = get_next; 271 stdp->std_getp = get_tailp; 272 stdp->std_get_count += count; 273 stdp->std_get_non_tcp_count += non_tcp_count; 274} 275 276static void 277sfxge_tx_qreap(struct sfxge_txq *txq) 278{ 279 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 280 281 txq->reaped = txq->completed; 282} 283 284static void 285sfxge_tx_qlist_post(struct sfxge_txq *txq) 286{ 287 unsigned int old_added; 288 unsigned int block_level; 289 unsigned int level; 290 int rc; 291 292 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 293 294 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0")); 295 KASSERT(txq->n_pend_desc <= txq->max_pkt_desc, 296 ("txq->n_pend_desc too large")); 297 KASSERT(!txq->blocked, ("txq->blocked")); 298 299 old_added = txq->added; 300 301 /* Post the fragment list. */ 302 rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc, 303 txq->reaped, &txq->added); 304 KASSERT(rc == 0, ("efx_tx_qdesc_post() failed")); 305 306 /* If efx_tx_qdesc_post() had to refragment, our information about 307 * buffers to free may be associated with the wrong 308 * descriptors. 309 */ 310 KASSERT(txq->added - old_added == txq->n_pend_desc, 311 ("efx_tx_qdesc_post() refragmented descriptors")); 312 313 level = txq->added - txq->reaped; 314 KASSERT(level <= txq->entries, ("overfilled TX queue")); 315 316 /* Clear the fragment list. */ 317 txq->n_pend_desc = 0; 318 319 /* 320 * Set the block level to ensure there is space to generate a 321 * large number of descriptors for TSO. 322 */ 323 block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc; 324 325 /* Have we reached the block level? */ 326 if (level < block_level) 327 return; 328 329 /* Reap, and check again */ 330 sfxge_tx_qreap(txq); 331 level = txq->added - txq->reaped; 332 if (level < block_level) 333 return; 334 335 txq->blocked = 1; 336 337 /* 338 * Avoid a race with completion interrupt handling that could leave 339 * the queue blocked. 340 */ 341 mb(); 342 sfxge_tx_qreap(txq); 343 level = txq->added - txq->reaped; 344 if (level < block_level) { 345 mb(); 346 txq->blocked = 0; 347 } 348} 349 350static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf) 351{ 352 bus_dmamap_t *used_map; 353 bus_dmamap_t map; 354 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG]; 355 unsigned int id; 356 struct sfxge_tx_mapping *stmp; 357 efx_desc_t *desc; 358 int n_dma_seg; 359 int rc; 360 int i; 361 int eop; 362 uint16_t hw_vlan_tci_prev; 363 int vlan_tagged; 364 365 KASSERT(!txq->blocked, ("txq->blocked")); 366 367#if SFXGE_TX_PARSE_EARLY 368 /* 369 * If software TSO is used, we still need to copy packet header, 370 * even if we have already parsed it early before enqueue. 371 */ 372 if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) && 373 (txq->tso_fw_assisted == 0)) 374 prefetch_read_many(mbuf->m_data); 375#else 376 /* 377 * Prefetch packet header since we need to parse it and extract 378 * IP ID, TCP sequence number and flags. 379 */ 380 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) 381 prefetch_read_many(mbuf->m_data); 382#endif 383 384 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) { 385 rc = EINTR; 386 goto reject; 387 } 388 389 /* Load the packet for DMA. */ 390 id = txq->added & txq->ptr_mask; 391 stmp = &txq->stmp[id]; 392 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, 393 mbuf, dma_seg, &n_dma_seg, 0); 394 if (rc == EFBIG) { 395 /* Try again. */ 396 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, 397 SFXGE_TX_MAPPING_MAX_SEG); 398 if (new_mbuf == NULL) 399 goto reject; 400 ++txq->collapses; 401 mbuf = new_mbuf; 402 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, 403 stmp->map, mbuf, 404 dma_seg, &n_dma_seg, 0); 405 } 406 if (rc != 0) 407 goto reject; 408 409 /* Make the packet visible to the hardware. */ 410 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); 411 412 used_map = &stmp->map; 413 414 hw_vlan_tci_prev = txq->hw_vlan_tci; 415 416 vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf); 417 if (vlan_tagged) { 418 sfxge_next_stmp(txq, &stmp); 419 } 420 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 421 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged); 422 if (rc < 0) 423 goto reject_mapped; 424 stmp = &txq->stmp[(rc - 1) & txq->ptr_mask]; 425 } else { 426 /* Add the mapping to the fragment list, and set flags 427 * for the buffer. 428 */ 429 430 i = 0; 431 for (;;) { 432 desc = &txq->pend_desc[i + vlan_tagged]; 433 eop = (i == n_dma_seg - 1); 434 efx_tx_qdesc_dma_create(txq->common, 435 dma_seg[i].ds_addr, 436 dma_seg[i].ds_len, 437 eop, 438 desc); 439 if (eop) 440 break; 441 i++; 442 sfxge_next_stmp(txq, &stmp); 443 } 444 txq->n_pend_desc = n_dma_seg + vlan_tagged; 445 } 446 447 /* 448 * If the mapping required more than one descriptor 449 * then we need to associate the DMA map with the last 450 * descriptor, not the first. 451 */ 452 if (used_map != &stmp->map) { 453 map = stmp->map; 454 stmp->map = *used_map; 455 *used_map = map; 456 } 457 458 stmp->u.mbuf = mbuf; 459 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; 460 461 /* Post the fragment list. */ 462 sfxge_tx_qlist_post(txq); 463 464 return (0); 465 466reject_mapped: 467 txq->hw_vlan_tci = hw_vlan_tci_prev; 468 bus_dmamap_unload(txq->packet_dma_tag, *used_map); 469reject: 470 /* Drop the packet on the floor. */ 471 m_freem(mbuf); 472 ++txq->drops; 473 474 return (rc); 475} 476 477/* 478 * Drain the deferred packet list into the transmit queue. 479 */ 480static void 481sfxge_tx_qdpl_drain(struct sfxge_txq *txq) 482{ 483 struct sfxge_softc *sc; 484 struct sfxge_tx_dpl *stdp; 485 struct mbuf *mbuf, *next; 486 unsigned int count; 487 unsigned int non_tcp_count; 488 unsigned int pushed; 489 int rc; 490 491 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 492 493 sc = txq->sc; 494 stdp = &txq->dpl; 495 pushed = txq->added; 496 497 if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) { 498 prefetch_read_many(sc->enp); 499 prefetch_read_many(txq->common); 500 } 501 502 mbuf = stdp->std_get; 503 count = stdp->std_get_count; 504 non_tcp_count = stdp->std_get_non_tcp_count; 505 506 if (count > stdp->std_get_hiwat) 507 stdp->std_get_hiwat = count; 508 509 while (count != 0) { 510 KASSERT(mbuf != NULL, ("mbuf == NULL")); 511 512 next = mbuf->m_nextpkt; 513 mbuf->m_nextpkt = NULL; 514 515 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ 516 517 if (next != NULL) 518 prefetch_read_many(next); 519 520 rc = sfxge_tx_queue_mbuf(txq, mbuf); 521 --count; 522 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); 523 mbuf = next; 524 if (rc != 0) 525 continue; 526 527 if (txq->blocked) 528 break; 529 530 /* Push the fragments to the hardware in batches. */ 531 if (txq->added - pushed >= SFXGE_TX_BATCH) { 532 efx_tx_qpush(txq->common, txq->added, pushed); 533 pushed = txq->added; 534 } 535 } 536 537 if (count == 0) { 538 KASSERT(mbuf == NULL, ("mbuf != NULL")); 539 KASSERT(non_tcp_count == 0, 540 ("inconsistent TCP/non-TCP detection")); 541 stdp->std_get = NULL; 542 stdp->std_get_count = 0; 543 stdp->std_get_non_tcp_count = 0; 544 stdp->std_getp = &stdp->std_get; 545 } else { 546 stdp->std_get = mbuf; 547 stdp->std_get_count = count; 548 stdp->std_get_non_tcp_count = non_tcp_count; 549 } 550 551 if (txq->added != pushed) 552 efx_tx_qpush(txq->common, txq->added, pushed); 553 554 KASSERT(txq->blocked || stdp->std_get_count == 0, 555 ("queue unblocked but count is non-zero")); 556} 557 558#define SFXGE_TX_QDPL_PENDING(_txq) ((_txq)->dpl.std_put != 0) 559 560/* 561 * Service the deferred packet list. 562 * 563 * NOTE: drops the txq mutex! 564 */ 565static void 566sfxge_tx_qdpl_service(struct sfxge_txq *txq) 567{ 568 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 569 570 do { 571 if (SFXGE_TX_QDPL_PENDING(txq)) 572 sfxge_tx_qdpl_swizzle(txq); 573 574 if (!txq->blocked) 575 sfxge_tx_qdpl_drain(txq); 576 577 SFXGE_TXQ_UNLOCK(txq); 578 } while (SFXGE_TX_QDPL_PENDING(txq) && 579 SFXGE_TXQ_TRYLOCK(txq)); 580} 581 582/* 583 * Put a packet on the deferred packet get-list. 584 */ 585static int 586sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf) 587{ 588 struct sfxge_tx_dpl *stdp; 589 590 stdp = &txq->dpl; 591 592 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 593 594 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 595 596 if (stdp->std_get_count >= stdp->std_get_max) { 597 txq->get_overflow++; 598 return (ENOBUFS); 599 } 600 if (sfxge_is_mbuf_non_tcp(mbuf)) { 601 if (stdp->std_get_non_tcp_count >= 602 stdp->std_get_non_tcp_max) { 603 txq->get_non_tcp_overflow++; 604 return (ENOBUFS); 605 } 606 stdp->std_get_non_tcp_count++; 607 } 608 609 *(stdp->std_getp) = mbuf; 610 stdp->std_getp = &mbuf->m_nextpkt; 611 stdp->std_get_count++; 612 613 return (0); 614} 615 616/* 617 * Put a packet on the deferred packet put-list. 618 * 619 * We overload the csum_data field in the mbuf to keep track of this length 620 * because there is no cheap alternative to avoid races. 621 */ 622static int 623sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf) 624{ 625 struct sfxge_tx_dpl *stdp; 626 volatile uintptr_t *putp; 627 uintptr_t old; 628 uintptr_t new; 629 unsigned int put_count; 630 631 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 632 633 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 634 635 stdp = &txq->dpl; 636 putp = &stdp->std_put; 637 new = (uintptr_t)mbuf; 638 639 do { 640 old = *putp; 641 if (old != 0) { 642 struct mbuf *mp = (struct mbuf *)old; 643 put_count = mp->m_pkthdr.csum_data; 644 } else 645 put_count = 0; 646 if (put_count >= stdp->std_put_max) { 647 atomic_add_long(&txq->put_overflow, 1); 648 return (ENOBUFS); 649 } 650 mbuf->m_pkthdr.csum_data = put_count + 1; 651 mbuf->m_nextpkt = (void *)old; 652 } while (atomic_cmpset_ptr(putp, old, new) == 0); 653 654 return (0); 655} 656 657/* 658 * Called from if_transmit - will try to grab the txq lock and enqueue to the 659 * put list if it succeeds, otherwise try to push onto the defer list if space. 660 */ 661static int 662sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) 663{ 664 int rc; 665 666 if (!SFXGE_LINK_UP(txq->sc)) { 667 atomic_add_long(&txq->netdown_drops, 1); 668 return (ENETDOWN); 669 } 670 671 /* 672 * Try to grab the txq lock. If we are able to get the lock, 673 * the packet will be appended to the "get list" of the deferred 674 * packet list. Otherwise, it will be pushed on the "put list". 675 */ 676 if (SFXGE_TXQ_TRYLOCK(txq)) { 677 /* First swizzle put-list to get-list to keep order */ 678 sfxge_tx_qdpl_swizzle(txq); 679 680 rc = sfxge_tx_qdpl_put_locked(txq, m); 681 682 /* Try to service the list. */ 683 sfxge_tx_qdpl_service(txq); 684 /* Lock has been dropped. */ 685 } else { 686 rc = sfxge_tx_qdpl_put_unlocked(txq, m); 687 688 /* 689 * Try to grab the lock again. 690 * 691 * If we are able to get the lock, we need to process 692 * the deferred packet list. If we are not able to get 693 * the lock, another thread is processing the list. 694 */ 695 if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) { 696 sfxge_tx_qdpl_service(txq); 697 /* Lock has been dropped. */ 698 } 699 } 700 701 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 702 703 return (rc); 704} 705 706static void 707sfxge_tx_qdpl_flush(struct sfxge_txq *txq) 708{ 709 struct sfxge_tx_dpl *stdp = &txq->dpl; 710 struct mbuf *mbuf, *next; 711 712 SFXGE_TXQ_LOCK(txq); 713 714 sfxge_tx_qdpl_swizzle(txq); 715 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { 716 next = mbuf->m_nextpkt; 717 m_freem(mbuf); 718 } 719 stdp->std_get = NULL; 720 stdp->std_get_count = 0; 721 stdp->std_get_non_tcp_count = 0; 722 stdp->std_getp = &stdp->std_get; 723 724 SFXGE_TXQ_UNLOCK(txq); 725} 726 727void 728sfxge_if_qflush(struct ifnet *ifp) 729{ 730 struct sfxge_softc *sc; 731 unsigned int i; 732 733 sc = ifp->if_softc; 734 735 for (i = 0; i < sc->txq_count; i++) 736 sfxge_tx_qdpl_flush(sc->txq[i]); 737} 738 739#if SFXGE_TX_PARSE_EARLY 740 741/* There is little space for user data in mbuf pkthdr, so we 742 * use l*hlen fields which are not used by the driver otherwise 743 * to store header offsets. 744 * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes. 745 */ 746 747 748#define TSO_MBUF_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[0]) 749/* We abuse l5hlen here because PH_loc can hold only 64 bits of data */ 750#define TSO_MBUF_FLAGS(_mbuf) ((_mbuf)->m_pkthdr.l5hlen) 751#define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1]) 752#define TSO_MBUF_SEQNUM(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1]) 753 754static void sfxge_parse_tx_packet(struct mbuf *mbuf) 755{ 756 struct ether_header *eh = mtod(mbuf, struct ether_header *); 757 const struct tcphdr *th; 758 struct tcphdr th_copy; 759 760 /* Find network protocol and header */ 761 TSO_MBUF_PROTO(mbuf) = eh->ether_type; 762 if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) { 763 struct ether_vlan_header *veh = 764 mtod(mbuf, struct ether_vlan_header *); 765 TSO_MBUF_PROTO(mbuf) = veh->evl_proto; 766 mbuf->m_pkthdr.l2hlen = sizeof(*veh); 767 } else { 768 mbuf->m_pkthdr.l2hlen = sizeof(*eh); 769 } 770 771 /* Find TCP header */ 772 if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) { 773 const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen); 774 775 KASSERT(iph->ip_p == IPPROTO_TCP, 776 ("TSO required on non-TCP packet")); 777 mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl; 778 TSO_MBUF_PACKETID(mbuf) = iph->ip_id; 779 } else { 780 KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6), 781 ("TSO required on non-IP packet")); 782 KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt == 783 IPPROTO_TCP, 784 ("TSO required on non-TCP packet")); 785 mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr); 786 TSO_MBUF_PACKETID(mbuf) = 0; 787 } 788 789 KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen, 790 ("network header is fragmented in mbuf")); 791 792 /* We need TCP header including flags (window is the next) */ 793 if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) { 794 m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy), 795 (caddr_t)&th_copy); 796 th = &th_copy; 797 } else { 798 th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen); 799 } 800 801 mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off; 802 TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq); 803 804 /* These flags must not be duplicated */ 805 /* 806 * RST should not be duplicated as well, but FreeBSD kernel 807 * generates TSO packets with RST flag. So, do not assert 808 * its absence. 809 */ 810 KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), 811 ("incompatible TCP flag 0x%x on TSO packet", 812 th->th_flags & (TH_URG | TH_SYN))); 813 TSO_MBUF_FLAGS(mbuf) = th->th_flags; 814} 815#endif 816 817/* 818 * TX start -- called by the stack. 819 */ 820int 821sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) 822{ 823 struct sfxge_softc *sc; 824 struct sfxge_txq *txq; 825 int rc; 826 827 sc = (struct sfxge_softc *)ifp->if_softc; 828 829 /* 830 * Transmit may be called when interface is up from the kernel 831 * point of view, but not yet up (in progress) from the driver 832 * point of view. I.e. link aggregation bring up. 833 * Transmit may be called when interface is up from the driver 834 * point of view, but already down from the kernel point of 835 * view. I.e. Rx when interface shutdown is in progress. 836 */ 837 KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP), 838 ("interface not up")); 839 840 /* Pick the desired transmit queue. */ 841 if (m->m_pkthdr.csum_flags & 842 (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) { 843 int index = 0; 844 845#ifdef RSS 846 uint32_t bucket_id; 847 848 /* 849 * Select a TX queue which matches the corresponding 850 * RX queue for the hash in order to assign both 851 * TX and RX parts of the flow to the same CPU 852 */ 853 if (rss_m2bucket(m, &bucket_id) == 0) 854 index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1)); 855#else 856 /* check if flowid is set */ 857 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 858 uint32_t hash = m->m_pkthdr.flowid; 859 uint32_t idx = hash % nitems(sc->rx_indir_table); 860 861 index = sc->rx_indir_table[idx]; 862 } 863#endif 864#if SFXGE_TX_PARSE_EARLY 865 if (m->m_pkthdr.csum_flags & CSUM_TSO) 866 sfxge_parse_tx_packet(m); 867#endif 868 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; 869 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { 870 txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; 871 } else { 872 txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; 873 } 874 875 rc = sfxge_tx_packet_add(txq, m); 876 if (rc != 0) 877 m_freem(m); 878 879 return (rc); 880} 881 882/* 883 * Software "TSO". Not quite as good as doing it in hardware, but 884 * still faster than segmenting in the stack. 885 */ 886 887struct sfxge_tso_state { 888 /* Output position */ 889 unsigned out_len; /* Remaining length in current segment */ 890 unsigned seqnum; /* Current sequence number */ 891 unsigned packet_space; /* Remaining space in current packet */ 892 unsigned segs_space; /* Remaining number of DMA segments 893 for the packet (FATSOv2 only) */ 894 895 /* Input position */ 896 uint64_t dma_addr; /* DMA address of current position */ 897 unsigned in_len; /* Remaining length in current mbuf */ 898 899 const struct mbuf *mbuf; /* Input mbuf (head of chain) */ 900 u_short protocol; /* Network protocol (after VLAN decap) */ 901 ssize_t nh_off; /* Offset of network header */ 902 ssize_t tcph_off; /* Offset of TCP header */ 903 unsigned header_len; /* Number of bytes of header */ 904 unsigned seg_size; /* TCP segment size */ 905 int fw_assisted; /* Use FW-assisted TSO */ 906 u_short packet_id; /* IPv4 packet ID from the original packet */ 907 uint8_t tcp_flags; /* TCP flags */ 908 efx_desc_t header_desc; /* Precomputed header descriptor for 909 * FW-assisted TSO */ 910}; 911 912#if !SFXGE_TX_PARSE_EARLY 913static const struct ip *tso_iph(const struct sfxge_tso_state *tso) 914{ 915 KASSERT(tso->protocol == htons(ETHERTYPE_IP), 916 ("tso_iph() in non-IPv4 state")); 917 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); 918} 919 920static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) 921{ 922 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 923 ("tso_ip6h() in non-IPv6 state")); 924 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); 925} 926 927static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) 928{ 929 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); 930} 931#endif 932 933 934/* Size of preallocated TSO header buffers. Larger blocks must be 935 * allocated from the heap. 936 */ 937#define TSOH_STD_SIZE 128 938 939/* At most half the descriptors in the queue at any time will refer to 940 * a TSO header buffer, since they must always be followed by a 941 * payload descriptor referring to an mbuf. 942 */ 943#define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) 944#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 945#define TSOH_PAGE_COUNT(_txq_entries) \ 946 howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE) 947 948static int tso_init(struct sfxge_txq *txq) 949{ 950 struct sfxge_softc *sc = txq->sc; 951 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); 952 int i, rc; 953 954 /* Allocate TSO header buffers */ 955 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), 956 M_SFXGE, M_WAITOK); 957 958 for (i = 0; i < tsoh_page_count; i++) { 959 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); 960 if (rc != 0) 961 goto fail; 962 } 963 964 return (0); 965 966fail: 967 while (i-- > 0) 968 sfxge_dma_free(&txq->tsoh_buffer[i]); 969 free(txq->tsoh_buffer, M_SFXGE); 970 txq->tsoh_buffer = NULL; 971 return (rc); 972} 973 974static void tso_fini(struct sfxge_txq *txq) 975{ 976 int i; 977 978 if (txq->tsoh_buffer != NULL) { 979 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) 980 sfxge_dma_free(&txq->tsoh_buffer[i]); 981 free(txq->tsoh_buffer, M_SFXGE); 982 } 983} 984 985static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso, 986 const bus_dma_segment_t *hdr_dma_seg, 987 struct mbuf *mbuf) 988{ 989 const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp); 990#if !SFXGE_TX_PARSE_EARLY 991 struct ether_header *eh = mtod(mbuf, struct ether_header *); 992 const struct tcphdr *th; 993 struct tcphdr th_copy; 994#endif 995 996 tso->fw_assisted = txq->tso_fw_assisted; 997 tso->mbuf = mbuf; 998 999 /* Find network protocol and header */ 1000#if !SFXGE_TX_PARSE_EARLY 1001 tso->protocol = eh->ether_type; 1002 if (tso->protocol == htons(ETHERTYPE_VLAN)) { 1003 struct ether_vlan_header *veh = 1004 mtod(mbuf, struct ether_vlan_header *); 1005 tso->protocol = veh->evl_proto; 1006 tso->nh_off = sizeof(*veh); 1007 } else { 1008 tso->nh_off = sizeof(*eh); 1009 } 1010#else 1011 tso->protocol = TSO_MBUF_PROTO(mbuf); 1012 tso->nh_off = mbuf->m_pkthdr.l2hlen; 1013 tso->tcph_off = mbuf->m_pkthdr.l3hlen; 1014 tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf)); 1015#endif 1016 1017#if !SFXGE_TX_PARSE_EARLY 1018 /* Find TCP header */ 1019 if (tso->protocol == htons(ETHERTYPE_IP)) { 1020 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, 1021 ("TSO required on non-TCP packet")); 1022 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; 1023 tso->packet_id = ntohs(tso_iph(tso)->ip_id); 1024 } else { 1025 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 1026 ("TSO required on non-IP packet")); 1027 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, 1028 ("TSO required on non-TCP packet")); 1029 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); 1030 tso->packet_id = 0; 1031 } 1032#endif 1033 1034 1035 if (tso->fw_assisted && 1036 __predict_false(tso->tcph_off > 1037 encp->enc_tx_tso_tcp_header_offset_limit)) { 1038 tso->fw_assisted = 0; 1039 } 1040 1041 1042#if !SFXGE_TX_PARSE_EARLY 1043 KASSERT(mbuf->m_len >= tso->tcph_off, 1044 ("network header is fragmented in mbuf")); 1045 /* We need TCP header including flags (window is the next) */ 1046 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) { 1047 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy), 1048 (caddr_t)&th_copy); 1049 th = &th_copy; 1050 } else { 1051 th = tso_tcph(tso); 1052 } 1053 tso->header_len = tso->tcph_off + 4 * th->th_off; 1054#else 1055 tso->header_len = mbuf->m_pkthdr.l4hlen; 1056#endif 1057 tso->seg_size = mbuf->m_pkthdr.tso_segsz; 1058 1059#if !SFXGE_TX_PARSE_EARLY 1060 tso->seqnum = ntohl(th->th_seq); 1061 1062 /* These flags must not be duplicated */ 1063 /* 1064 * RST should not be duplicated as well, but FreeBSD kernel 1065 * generates TSO packets with RST flag. So, do not assert 1066 * its absence. 1067 */ 1068 KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), 1069 ("incompatible TCP flag 0x%x on TSO packet", 1070 th->th_flags & (TH_URG | TH_SYN))); 1071 tso->tcp_flags = th->th_flags; 1072#else 1073 tso->seqnum = TSO_MBUF_SEQNUM(mbuf); 1074 tso->tcp_flags = TSO_MBUF_FLAGS(mbuf); 1075#endif 1076 1077 tso->out_len = mbuf->m_pkthdr.len - tso->header_len; 1078 1079 if (tso->fw_assisted) { 1080 if (hdr_dma_seg->ds_len >= tso->header_len) 1081 efx_tx_qdesc_dma_create(txq->common, 1082 hdr_dma_seg->ds_addr, 1083 tso->header_len, 1084 B_FALSE, 1085 &tso->header_desc); 1086 else 1087 tso->fw_assisted = 0; 1088 } 1089} 1090 1091/* 1092 * tso_fill_packet_with_fragment - form descriptors for the current fragment 1093 * 1094 * Form descriptors for the current fragment, until we reach the end 1095 * of fragment or end-of-packet. Return 0 on success, 1 if not enough 1096 * space. 1097 */ 1098static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, 1099 struct sfxge_tso_state *tso) 1100{ 1101 efx_desc_t *desc; 1102 int n; 1103 uint64_t dma_addr = tso->dma_addr; 1104 boolean_t eop; 1105 1106 if (tso->in_len == 0 || tso->packet_space == 0) 1107 return; 1108 1109 KASSERT(tso->in_len > 0, ("TSO input length went negative")); 1110 KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); 1111 1112 if (tso->fw_assisted & SFXGE_FATSOV2) { 1113 n = tso->in_len; 1114 tso->out_len -= n; 1115 tso->seqnum += n; 1116 tso->in_len = 0; 1117 if (n < tso->packet_space) { 1118 tso->packet_space -= n; 1119 tso->segs_space--; 1120 } else { 1121 tso->packet_space = tso->seg_size - 1122 (n - tso->packet_space) % tso->seg_size; 1123 tso->segs_space = 1124 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 - 1125 (tso->packet_space != tso->seg_size); 1126 } 1127 } else { 1128 n = min(tso->in_len, tso->packet_space); 1129 tso->packet_space -= n; 1130 tso->out_len -= n; 1131 tso->dma_addr += n; 1132 tso->in_len -= n; 1133 } 1134 1135 /* 1136 * It is OK to use binary OR below to avoid extra branching 1137 * since all conditions may always be checked. 1138 */ 1139 eop = (tso->out_len == 0) | (tso->packet_space == 0) | 1140 (tso->segs_space == 0); 1141 1142 desc = &txq->pend_desc[txq->n_pend_desc++]; 1143 efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc); 1144} 1145 1146/* Callback from bus_dmamap_load() for long TSO headers. */ 1147static void tso_map_long_header(void *dma_addr_ret, 1148 bus_dma_segment_t *segs, int nseg, 1149 int error) 1150{ 1151 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && 1152 __predict_true(nseg == 1)) ? 1153 segs->ds_addr : 0); 1154} 1155 1156/* 1157 * tso_start_new_packet - generate a new header and prepare for the new packet 1158 * 1159 * Generate a new header and prepare for the new packet. Return 0 on 1160 * success, or an error code if failed to alloc header. 1161 */ 1162static int tso_start_new_packet(struct sfxge_txq *txq, 1163 struct sfxge_tso_state *tso, 1164 unsigned int *idp) 1165{ 1166 unsigned int id = *idp; 1167 struct tcphdr *tsoh_th; 1168 unsigned ip_length; 1169 caddr_t header; 1170 uint64_t dma_addr; 1171 bus_dmamap_t map; 1172 efx_desc_t *desc; 1173 int rc; 1174 1175 if (tso->fw_assisted) { 1176 if (tso->fw_assisted & SFXGE_FATSOV2) { 1177 /* Add 2 FATSOv2 option descriptors */ 1178 desc = &txq->pend_desc[txq->n_pend_desc]; 1179 efx_tx_qdesc_tso2_create(txq->common, 1180 tso->packet_id, 1181 tso->seqnum, 1182 tso->seg_size, 1183 desc, 1184 EFX_TX_FATSOV2_OPT_NDESCS); 1185 desc += EFX_TX_FATSOV2_OPT_NDESCS; 1186 txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS; 1187 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1188 id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask; 1189 1190 tso->segs_space = 1191 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1; 1192 } else { 1193 uint8_t tcp_flags = tso->tcp_flags; 1194 1195 if (tso->out_len > tso->seg_size) 1196 tcp_flags &= ~(TH_FIN | TH_PUSH); 1197 1198 /* Add FATSOv1 option descriptor */ 1199 desc = &txq->pend_desc[txq->n_pend_desc++]; 1200 efx_tx_qdesc_tso_create(txq->common, 1201 tso->packet_id, 1202 tso->seqnum, 1203 tcp_flags, 1204 desc++); 1205 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1206 id = (id + 1) & txq->ptr_mask; 1207 1208 tso->seqnum += tso->seg_size; 1209 tso->segs_space = UINT_MAX; 1210 } 1211 1212 /* Header DMA descriptor */ 1213 *desc = tso->header_desc; 1214 txq->n_pend_desc++; 1215 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1216 id = (id + 1) & txq->ptr_mask; 1217 } else { 1218 /* Allocate a DMA-mapped header buffer. */ 1219 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { 1220 unsigned int page_index = (id / 2) / TSOH_PER_PAGE; 1221 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; 1222 1223 header = (txq->tsoh_buffer[page_index].esm_base + 1224 buf_index * TSOH_STD_SIZE); 1225 dma_addr = (txq->tsoh_buffer[page_index].esm_addr + 1226 buf_index * TSOH_STD_SIZE); 1227 map = txq->tsoh_buffer[page_index].esm_map; 1228 1229 KASSERT(txq->stmp[id].flags == 0, 1230 ("stmp flags are not 0")); 1231 } else { 1232 struct sfxge_tx_mapping *stmp = &txq->stmp[id]; 1233 1234 /* We cannot use bus_dmamem_alloc() as that may sleep */ 1235 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); 1236 if (__predict_false(!header)) 1237 return (ENOMEM); 1238 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, 1239 header, tso->header_len, 1240 tso_map_long_header, &dma_addr, 1241 BUS_DMA_NOWAIT); 1242 if (__predict_false(dma_addr == 0)) { 1243 if (rc == 0) { 1244 /* Succeeded but got >1 segment */ 1245 bus_dmamap_unload(txq->packet_dma_tag, 1246 stmp->map); 1247 rc = EINVAL; 1248 } 1249 free(header, M_SFXGE); 1250 return (rc); 1251 } 1252 map = stmp->map; 1253 1254 txq->tso_long_headers++; 1255 stmp->u.heap_buf = header; 1256 stmp->flags = TX_BUF_UNMAP; 1257 } 1258 1259 tsoh_th = (struct tcphdr *)(header + tso->tcph_off); 1260 1261 /* Copy and update the headers. */ 1262 m_copydata(tso->mbuf, 0, tso->header_len, header); 1263 1264 tsoh_th->th_seq = htonl(tso->seqnum); 1265 tso->seqnum += tso->seg_size; 1266 if (tso->out_len > tso->seg_size) { 1267 /* This packet will not finish the TSO burst. */ 1268 ip_length = tso->header_len - tso->nh_off + tso->seg_size; 1269 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); 1270 } else { 1271 /* This packet will be the last in the TSO burst. */ 1272 ip_length = tso->header_len - tso->nh_off + tso->out_len; 1273 } 1274 1275 if (tso->protocol == htons(ETHERTYPE_IP)) { 1276 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); 1277 tsoh_iph->ip_len = htons(ip_length); 1278 /* XXX We should increment ip_id, but FreeBSD doesn't 1279 * currently allocate extra IDs for multiple segments. 1280 */ 1281 } else { 1282 struct ip6_hdr *tsoh_iph = 1283 (struct ip6_hdr *)(header + tso->nh_off); 1284 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); 1285 } 1286 1287 /* Make the header visible to the hardware. */ 1288 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); 1289 1290 /* Form a descriptor for this header. */ 1291 desc = &txq->pend_desc[txq->n_pend_desc++]; 1292 efx_tx_qdesc_dma_create(txq->common, 1293 dma_addr, 1294 tso->header_len, 1295 0, 1296 desc); 1297 id = (id + 1) & txq->ptr_mask; 1298 1299 tso->segs_space = UINT_MAX; 1300 } 1301 tso->packet_space = tso->seg_size; 1302 txq->tso_packets++; 1303 *idp = id; 1304 1305 return (0); 1306} 1307 1308static int 1309sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 1310 const bus_dma_segment_t *dma_seg, int n_dma_seg, 1311 int vlan_tagged) 1312{ 1313 struct sfxge_tso_state tso; 1314 unsigned int id; 1315 unsigned skipped = 0; 1316 1317 tso_start(txq, &tso, dma_seg, mbuf); 1318 1319 while (dma_seg->ds_len + skipped <= tso.header_len) { 1320 skipped += dma_seg->ds_len; 1321 --n_dma_seg; 1322 KASSERT(n_dma_seg, ("no payload found in TSO packet")); 1323 ++dma_seg; 1324 } 1325 tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); 1326 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); 1327 1328 id = (txq->added + vlan_tagged) & txq->ptr_mask; 1329 if (__predict_false(tso_start_new_packet(txq, &tso, &id))) 1330 return (-1); 1331 1332 while (1) { 1333 tso_fill_packet_with_fragment(txq, &tso); 1334 /* Exactly one DMA descriptor is added */ 1335 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1336 id = (id + 1) & txq->ptr_mask; 1337 1338 /* Move onto the next fragment? */ 1339 if (tso.in_len == 0) { 1340 --n_dma_seg; 1341 if (n_dma_seg == 0) 1342 break; 1343 ++dma_seg; 1344 tso.in_len = dma_seg->ds_len; 1345 tso.dma_addr = dma_seg->ds_addr; 1346 } 1347 1348 /* End of packet? */ 1349 if ((tso.packet_space == 0) | (tso.segs_space == 0)) { 1350 unsigned int n_fatso_opt_desc = 1351 (tso.fw_assisted & SFXGE_FATSOV2) ? 1352 EFX_TX_FATSOV2_OPT_NDESCS : 1353 (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0; 1354 1355 /* If the queue is now full due to tiny MSS, 1356 * or we can't create another header, discard 1357 * the remainder of the input mbuf but do not 1358 * roll back the work we have done. 1359 */ 1360 if (txq->n_pend_desc + n_fatso_opt_desc + 1361 1 /* header */ + n_dma_seg > txq->max_pkt_desc) { 1362 txq->tso_pdrop_too_many++; 1363 break; 1364 } 1365 if (__predict_false(tso_start_new_packet(txq, &tso, 1366 &id))) { 1367 txq->tso_pdrop_no_rsrc++; 1368 break; 1369 } 1370 } 1371 } 1372 1373 txq->tso_bursts++; 1374 return (id); 1375} 1376 1377static void 1378sfxge_tx_qunblock(struct sfxge_txq *txq) 1379{ 1380 struct sfxge_softc *sc; 1381 struct sfxge_evq *evq; 1382 1383 sc = txq->sc; 1384 evq = sc->evq[txq->evq_index]; 1385 1386 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 1387 1388 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) 1389 return; 1390 1391 SFXGE_TXQ_LOCK(txq); 1392 1393 if (txq->blocked) { 1394 unsigned int level; 1395 1396 level = txq->added - txq->completed; 1397 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) { 1398 /* reaped must be in sync with blocked */ 1399 sfxge_tx_qreap(txq); 1400 txq->blocked = 0; 1401 } 1402 } 1403 1404 sfxge_tx_qdpl_service(txq); 1405 /* note: lock has been dropped */ 1406} 1407 1408void 1409sfxge_tx_qflush_done(struct sfxge_txq *txq) 1410{ 1411 1412 txq->flush_state = SFXGE_FLUSH_DONE; 1413} 1414 1415static void 1416sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) 1417{ 1418 struct sfxge_txq *txq; 1419 struct sfxge_evq *evq; 1420 unsigned int count; 1421 1422 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1423 1424 txq = sc->txq[index]; 1425 evq = sc->evq[txq->evq_index]; 1426 1427 SFXGE_EVQ_LOCK(evq); 1428 SFXGE_TXQ_LOCK(txq); 1429 1430 KASSERT(txq->init_state == SFXGE_TXQ_STARTED, 1431 ("txq->init_state != SFXGE_TXQ_STARTED")); 1432 1433 txq->init_state = SFXGE_TXQ_INITIALIZED; 1434 1435 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1436 txq->flush_state = SFXGE_FLUSH_PENDING; 1437 1438 SFXGE_EVQ_UNLOCK(evq); 1439 SFXGE_TXQ_UNLOCK(txq); 1440 1441 /* Flush the transmit queue. */ 1442 if (efx_tx_qflush(txq->common) != 0) { 1443 log(LOG_ERR, "%s: Flushing Tx queue %u failed\n", 1444 device_get_nameunit(sc->dev), index); 1445 txq->flush_state = SFXGE_FLUSH_DONE; 1446 } else { 1447 count = 0; 1448 do { 1449 /* Spin for 100ms. */ 1450 DELAY(100000); 1451 if (txq->flush_state != SFXGE_FLUSH_PENDING) 1452 break; 1453 } while (++count < 20); 1454 } 1455 SFXGE_EVQ_LOCK(evq); 1456 SFXGE_TXQ_LOCK(txq); 1457 1458 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, 1459 ("txq->flush_state == SFXGE_FLUSH_FAILED")); 1460 1461 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1462 /* Flush timeout */ 1463 log(LOG_ERR, "%s: Cannot flush Tx queue %u\n", 1464 device_get_nameunit(sc->dev), index); 1465 txq->flush_state = SFXGE_FLUSH_DONE; 1466 } 1467 } 1468 1469 txq->blocked = 0; 1470 txq->pending = txq->added; 1471 1472 sfxge_tx_qcomplete(txq, evq); 1473 KASSERT(txq->completed == txq->added, 1474 ("txq->completed != txq->added")); 1475 1476 sfxge_tx_qreap(txq); 1477 KASSERT(txq->reaped == txq->completed, 1478 ("txq->reaped != txq->completed")); 1479 1480 txq->added = 0; 1481 txq->pending = 0; 1482 txq->completed = 0; 1483 txq->reaped = 0; 1484 1485 /* Destroy the common code transmit queue. */ 1486 efx_tx_qdestroy(txq->common); 1487 txq->common = NULL; 1488 1489 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1490 EFX_TXQ_NBUFS(sc->txq_entries)); 1491 1492 SFXGE_EVQ_UNLOCK(evq); 1493 SFXGE_TXQ_UNLOCK(txq); 1494} 1495 1496/* 1497 * Estimate maximum number of Tx descriptors required for TSO packet. 1498 * With minimum MSS and maximum mbuf length we might need more (even 1499 * than a ring-ful of descriptors), but this should not happen in 1500 * practice except due to deliberate attack. In that case we will 1501 * truncate the output at a packet boundary. 1502 */ 1503static unsigned int 1504sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type, 1505 unsigned int tso_fw_assisted) 1506{ 1507 /* One descriptor for every input fragment */ 1508 unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG; 1509 unsigned int sw_tso_max_descs; 1510 unsigned int fa_tso_v1_max_descs = 0; 1511 unsigned int fa_tso_v2_max_descs = 0; 1512 1513 /* VLAN tagging Tx option descriptor may be required */ 1514 if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled) 1515 max_descs++; 1516 1517 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) { 1518 /* 1519 * Plus header and payload descriptor for each output segment. 1520 * Minus one since header fragment is already counted. 1521 * Even if FATSO is used, we should be ready to fallback 1522 * to do it in the driver. 1523 */ 1524 sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1; 1525 1526 /* FW assisted TSOv1 requires one more descriptor per segment 1527 * in comparison to SW TSO */ 1528 if (tso_fw_assisted & SFXGE_FATSOV1) 1529 fa_tso_v1_max_descs = 1530 sw_tso_max_descs + SFXGE_TSO_MAX_SEGS; 1531 1532 /* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra 1533 * descriptors per superframe limited by number of DMA fetches 1534 * per packet. The first packet header is already counted. 1535 */ 1536 if (tso_fw_assisted & SFXGE_FATSOV2) { 1537 fa_tso_v2_max_descs = 1538 howmany(SFXGE_TX_MAPPING_MAX_SEG, 1539 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) * 1540 (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1; 1541 } 1542 1543 max_descs += MAX(sw_tso_max_descs, 1544 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs)); 1545 } 1546 1547 return (max_descs); 1548} 1549 1550static int 1551sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) 1552{ 1553 struct sfxge_txq *txq; 1554 efsys_mem_t *esmp; 1555 uint16_t flags; 1556 unsigned int tso_fw_assisted; 1557 struct sfxge_evq *evq; 1558 unsigned int desc_index; 1559 int rc; 1560 1561 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1562 1563 txq = sc->txq[index]; 1564 esmp = &txq->mem; 1565 evq = sc->evq[txq->evq_index]; 1566 1567 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1568 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1569 KASSERT(evq->init_state == SFXGE_EVQ_STARTED, 1570 ("evq->init_state != SFXGE_EVQ_STARTED")); 1571 1572 /* Program the buffer table. */ 1573 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, 1574 EFX_TXQ_NBUFS(sc->txq_entries))) != 0) 1575 return (rc); 1576 1577 /* Determine the kind of queue we are creating. */ 1578 tso_fw_assisted = 0; 1579 switch (txq->type) { 1580 case SFXGE_TXQ_NON_CKSUM: 1581 flags = 0; 1582 break; 1583 case SFXGE_TXQ_IP_CKSUM: 1584 flags = EFX_TXQ_CKSUM_IPV4; 1585 break; 1586 case SFXGE_TXQ_IP_TCP_UDP_CKSUM: 1587 flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP; 1588 tso_fw_assisted = sc->tso_fw_assisted; 1589 if (tso_fw_assisted & SFXGE_FATSOV2) 1590 flags |= EFX_TXQ_FATSOV2; 1591 break; 1592 default: 1593 KASSERT(0, ("Impossible TX queue")); 1594 flags = 0; 1595 break; 1596 } 1597 1598 /* Create the common code transmit queue. */ 1599 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1600 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1601 &txq->common, &desc_index)) != 0) { 1602 /* Retry if no FATSOv2 resources, otherwise fail */ 1603 if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2)) 1604 goto fail; 1605 1606 /* Looks like all FATSOv2 contexts are used */ 1607 flags &= ~EFX_TXQ_FATSOV2; 1608 tso_fw_assisted &= ~SFXGE_FATSOV2; 1609 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1610 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1611 &txq->common, &desc_index)) != 0) 1612 goto fail; 1613 } 1614 1615 /* Initialise queue descriptor indexes */ 1616 txq->added = txq->pending = txq->completed = txq->reaped = desc_index; 1617 1618 SFXGE_TXQ_LOCK(txq); 1619 1620 /* Enable the transmit queue. */ 1621 efx_tx_qenable(txq->common); 1622 1623 txq->init_state = SFXGE_TXQ_STARTED; 1624 txq->flush_state = SFXGE_FLUSH_REQUIRED; 1625 txq->tso_fw_assisted = tso_fw_assisted; 1626 1627 txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type, 1628 tso_fw_assisted); 1629 1630 txq->hw_vlan_tci = 0; 1631 1632 SFXGE_TXQ_UNLOCK(txq); 1633 1634 return (0); 1635 1636fail: 1637 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1638 EFX_TXQ_NBUFS(sc->txq_entries)); 1639 return (rc); 1640} 1641 1642void 1643sfxge_tx_stop(struct sfxge_softc *sc) 1644{ 1645 int index; 1646 1647 index = sc->txq_count; 1648 while (--index >= 0) 1649 sfxge_tx_qstop(sc, index); 1650 1651 /* Tear down the transmit module */ 1652 efx_tx_fini(sc->enp); 1653} 1654 1655int 1656sfxge_tx_start(struct sfxge_softc *sc) 1657{ 1658 int index; 1659 int rc; 1660 1661 /* Initialize the common code transmit module. */ 1662 if ((rc = efx_tx_init(sc->enp)) != 0) 1663 return (rc); 1664 1665 for (index = 0; index < sc->txq_count; index++) { 1666 if ((rc = sfxge_tx_qstart(sc, index)) != 0) 1667 goto fail; 1668 } 1669 1670 return (0); 1671 1672fail: 1673 while (--index >= 0) 1674 sfxge_tx_qstop(sc, index); 1675 1676 efx_tx_fini(sc->enp); 1677 1678 return (rc); 1679} 1680 1681static int 1682sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node) 1683{ 1684 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev); 1685 struct sysctl_oid *stat_node; 1686 unsigned int id; 1687 1688 stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1689 "stats", CTLFLAG_RD, NULL, 1690 "Tx queue statistics"); 1691 if (stat_node == NULL) 1692 return (ENOMEM); 1693 1694 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1695 SYSCTL_ADD_ULONG( 1696 ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO, 1697 sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS, 1698 (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset), 1699 ""); 1700 } 1701 1702 return (0); 1703} 1704 1705/** 1706 * Destroy a transmit queue. 1707 */ 1708static void 1709sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) 1710{ 1711 struct sfxge_txq *txq; 1712 unsigned int nmaps; 1713 1714 txq = sc->txq[index]; 1715 1716 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1717 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1718 1719 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) 1720 tso_fini(txq); 1721 1722 /* Free the context arrays. */ 1723 free(txq->pend_desc, M_SFXGE); 1724 nmaps = sc->txq_entries; 1725 while (nmaps-- != 0) 1726 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1727 free(txq->stmp, M_SFXGE); 1728 1729 /* Release DMA memory mapping. */ 1730 sfxge_dma_free(&txq->mem); 1731 1732 sc->txq[index] = NULL; 1733 1734 SFXGE_TXQ_LOCK_DESTROY(txq); 1735 1736 free(txq, M_SFXGE); 1737} 1738 1739static int 1740sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, 1741 enum sfxge_txq_type type, unsigned int evq_index) 1742{ 1743 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); 1744 char name[16]; 1745 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1746 struct sysctl_oid *txq_node; 1747 struct sfxge_txq *txq; 1748 struct sfxge_evq *evq; 1749 struct sfxge_tx_dpl *stdp; 1750 struct sysctl_oid *dpl_node; 1751 efsys_mem_t *esmp; 1752 unsigned int nmaps; 1753 int rc; 1754 1755 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); 1756 txq->sc = sc; 1757 txq->entries = sc->txq_entries; 1758 txq->ptr_mask = txq->entries - 1; 1759 1760 sc->txq[txq_index] = txq; 1761 esmp = &txq->mem; 1762 1763 evq = sc->evq[evq_index]; 1764 1765 /* Allocate and zero DMA space for the descriptor ring. */ 1766 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) 1767 return (rc); 1768 1769 /* Allocate buffer table entries. */ 1770 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), 1771 &txq->buf_base_id); 1772 1773 /* Create a DMA tag for packet mappings. */ 1774 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 1775 encp->enc_tx_dma_desc_boundary, 1776 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, 1777 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 1778 encp->enc_tx_dma_desc_size_max, 0, NULL, NULL, 1779 &txq->packet_dma_tag) != 0) { 1780 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); 1781 rc = ENOMEM; 1782 goto fail; 1783 } 1784 1785 /* Allocate pending descriptor array for batching writes. */ 1786 txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries, 1787 M_SFXGE, M_ZERO | M_WAITOK); 1788 1789 /* Allocate and initialise mbuf DMA mapping array. */ 1790 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, 1791 M_SFXGE, M_ZERO | M_WAITOK); 1792 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { 1793 rc = bus_dmamap_create(txq->packet_dma_tag, 0, 1794 &txq->stmp[nmaps].map); 1795 if (rc != 0) 1796 goto fail2; 1797 } 1798 1799 snprintf(name, sizeof(name), "%u", txq_index); 1800 txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node), 1801 OID_AUTO, name, CTLFLAG_RD, NULL, ""); 1802 if (txq_node == NULL) { 1803 rc = ENOMEM; 1804 goto fail_txq_node; 1805 } 1806 1807 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && 1808 (rc = tso_init(txq)) != 0) 1809 goto fail3; 1810 1811 /* Initialize the deferred packet list. */ 1812 stdp = &txq->dpl; 1813 stdp->std_put_max = sfxge_tx_dpl_put_max; 1814 stdp->std_get_max = sfxge_tx_dpl_get_max; 1815 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; 1816 stdp->std_getp = &stdp->std_get; 1817 1818 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index); 1819 1820 dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1821 "dpl", CTLFLAG_RD, NULL, 1822 "Deferred packet list statistics"); 1823 if (dpl_node == NULL) { 1824 rc = ENOMEM; 1825 goto fail_dpl_node; 1826 } 1827 1828 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1829 "get_count", CTLFLAG_RD | CTLFLAG_STATS, 1830 &stdp->std_get_count, 0, ""); 1831 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1832 "get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, 1833 &stdp->std_get_non_tcp_count, 0, ""); 1834 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1835 "get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1836 &stdp->std_get_hiwat, 0, ""); 1837 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1838 "put_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1839 &stdp->std_put_hiwat, 0, ""); 1840 1841 rc = sfxge_txq_stat_init(txq, txq_node); 1842 if (rc != 0) 1843 goto fail_txq_stat_init; 1844 1845 txq->type = type; 1846 txq->evq_index = evq_index; 1847 txq->init_state = SFXGE_TXQ_INITIALIZED; 1848 1849 return (0); 1850 1851fail_txq_stat_init: 1852fail_dpl_node: 1853fail3: 1854fail_txq_node: 1855 free(txq->pend_desc, M_SFXGE); 1856fail2: 1857 while (nmaps-- != 0) 1858 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1859 free(txq->stmp, M_SFXGE); 1860 bus_dma_tag_destroy(txq->packet_dma_tag); 1861 1862fail: 1863 sfxge_dma_free(esmp); 1864 1865 return (rc); 1866} 1867 1868static int 1869sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS) 1870{ 1871 struct sfxge_softc *sc = arg1; 1872 unsigned int id = arg2; 1873 unsigned long sum; 1874 unsigned int index; 1875 1876 /* Sum across all TX queues */ 1877 sum = 0; 1878 for (index = 0; index < sc->txq_count; index++) 1879 sum += *(unsigned long *)((caddr_t)sc->txq[index] + 1880 sfxge_tx_stats[id].offset); 1881 1882 return (SYSCTL_OUT(req, &sum, sizeof(sum))); 1883} 1884 1885static void 1886sfxge_tx_stat_init(struct sfxge_softc *sc) 1887{ 1888 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1889 struct sysctl_oid_list *stat_list; 1890 unsigned int id; 1891 1892 stat_list = SYSCTL_CHILDREN(sc->stats_node); 1893 1894 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1895 SYSCTL_ADD_PROC( 1896 ctx, stat_list, 1897 OID_AUTO, sfxge_tx_stats[id].name, 1898 CTLTYPE_ULONG|CTLFLAG_RD, 1899 sc, id, sfxge_tx_stat_handler, "LU", 1900 ""); 1901 } 1902} 1903 1904uint64_t 1905sfxge_tx_get_drops(struct sfxge_softc *sc) 1906{ 1907 unsigned int index; 1908 uint64_t drops = 0; 1909 struct sfxge_txq *txq; 1910 1911 /* Sum across all TX queues */ 1912 for (index = 0; index < sc->txq_count; index++) { 1913 txq = sc->txq[index]; 1914 /* 1915 * In theory, txq->put_overflow and txq->netdown_drops 1916 * should use atomic operation and other should be 1917 * obtained under txq lock, but it is just statistics. 1918 */ 1919 drops += txq->drops + txq->get_overflow + 1920 txq->get_non_tcp_overflow + 1921 txq->put_overflow + txq->netdown_drops + 1922 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc; 1923 } 1924 return (drops); 1925} 1926 1927void 1928sfxge_tx_fini(struct sfxge_softc *sc) 1929{ 1930 int index; 1931 1932 index = sc->txq_count; 1933 while (--index >= 0) 1934 sfxge_tx_qfini(sc, index); 1935 1936 sc->txq_count = 0; 1937} 1938 1939 1940int 1941sfxge_tx_init(struct sfxge_softc *sc) 1942{ 1943 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); 1944 struct sfxge_intr *intr; 1945 int index; 1946 int rc; 1947 1948 intr = &sc->intr; 1949 1950 KASSERT(intr->state == SFXGE_INTR_INITIALIZED, 1951 ("intr->state != SFXGE_INTR_INITIALIZED")); 1952 1953 if (sfxge_tx_dpl_get_max <= 0) { 1954 log(LOG_ERR, "%s=%d must be greater than 0", 1955 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); 1956 rc = EINVAL; 1957 goto fail_tx_dpl_get_max; 1958 } 1959 if (sfxge_tx_dpl_get_non_tcp_max <= 0) { 1960 log(LOG_ERR, "%s=%d must be greater than 0", 1961 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, 1962 sfxge_tx_dpl_get_non_tcp_max); 1963 rc = EINVAL; 1964 goto fail_tx_dpl_get_non_tcp_max; 1965 } 1966 if (sfxge_tx_dpl_put_max < 0) { 1967 log(LOG_ERR, "%s=%d must be greater or equal to 0", 1968 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); 1969 rc = EINVAL; 1970 goto fail_tx_dpl_put_max; 1971 } 1972 1973 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc; 1974 1975 sc->tso_fw_assisted = sfxge_tso_fw_assisted; 1976 if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) || 1977 (!encp->enc_fw_assisted_tso_enabled)) 1978 sc->tso_fw_assisted &= ~SFXGE_FATSOV1; 1979 if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) || 1980 (!encp->enc_fw_assisted_tso_v2_enabled)) 1981 sc->tso_fw_assisted &= ~SFXGE_FATSOV2; 1982 1983 sc->txqs_node = SYSCTL_ADD_NODE( 1984 device_get_sysctl_ctx(sc->dev), 1985 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), 1986 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); 1987 if (sc->txqs_node == NULL) { 1988 rc = ENOMEM; 1989 goto fail_txq_node; 1990 } 1991 1992 /* Initialize the transmit queues */ 1993 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, 1994 SFXGE_TXQ_NON_CKSUM, 0)) != 0) 1995 goto fail; 1996 1997 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, 1998 SFXGE_TXQ_IP_CKSUM, 0)) != 0) 1999 goto fail2; 2000 2001 for (index = 0; 2002 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1; 2003 index++) { 2004 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index, 2005 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) 2006 goto fail3; 2007 } 2008 2009 sfxge_tx_stat_init(sc); 2010 2011 return (0); 2012 2013fail3: 2014 while (--index >= 0) 2015 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 2016 2017 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 2018 2019fail2: 2020 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 2021 2022fail: 2023fail_txq_node: 2024 sc->txq_count = 0; 2025fail_tx_dpl_put_max: 2026fail_tx_dpl_get_non_tcp_max: 2027fail_tx_dpl_get_max: 2028 return (rc); 2029} 2030