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("$FreeBSD$"); 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) 84 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT; 85 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max); 86 SYSCTL_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) 92 static int sfxge_tx_dpl_get_non_tcp_max = 93 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT; 94 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max); 95 SYSCTL_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) 100 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT; 101 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max); 102 SYSCTL_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) 107 static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2); 108 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted); 109 SYSCTL_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"); 112 113 114 static 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. */ 135 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq); 136 static void sfxge_tx_qlist_post(struct sfxge_txq *txq); 137 static void sfxge_tx_qunblock(struct sfxge_txq *txq); 138 static 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); 141 142 static int 143 sfxge_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 } 159 160 static inline void 161 sfxge_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 } 170 171 172 void 173 sfxge_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 } 212 213 static unsigned int 214 sfxge_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 */ 225 static void 226 sfxge_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 } 275 276 static void 277 sfxge_tx_qreap(struct sfxge_txq *txq) 278 { 279 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 280 281 txq->reaped = txq->completed; 282 } 283 284 static void 285 sfxge_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 } 349 350 static 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 int vlan_tagged; 363 364 KASSERT(!txq->blocked, ("txq->blocked")); 365 366 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) 367 prefetch_read_many(mbuf->m_data); 368 369 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) { 370 rc = EINTR; 371 goto reject; 372 } 373 374 /* Load the packet for DMA. */ 375 id = txq->added & txq->ptr_mask; 376 stmp = &txq->stmp[id]; 377 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, 378 mbuf, dma_seg, &n_dma_seg, 0); 379 if (rc == EFBIG) { 380 /* Try again. */ 381 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, 382 SFXGE_TX_MAPPING_MAX_SEG); 383 if (new_mbuf == NULL) 384 goto reject; 385 ++txq->collapses; 386 mbuf = new_mbuf; 387 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, 388 stmp->map, mbuf, 389 dma_seg, &n_dma_seg, 0); 390 } 391 if (rc != 0) 392 goto reject; 393 394 /* Make the packet visible to the hardware. */ 395 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); 396 397 used_map = &stmp->map; 398 399 vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf); 400 if (vlan_tagged) { 401 sfxge_next_stmp(txq, &stmp); 402 } 403 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 404 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged); 405 if (rc < 0) 406 goto reject_mapped; 407 stmp = &txq->stmp[(rc - 1) & txq->ptr_mask]; 408 } else { 409 /* Add the mapping to the fragment list, and set flags 410 * for the buffer. 411 */ 412 413 i = 0; 414 for (;;) { 415 desc = &txq->pend_desc[i + vlan_tagged]; 416 eop = (i == n_dma_seg - 1); 417 efx_tx_qdesc_dma_create(txq->common, 418 dma_seg[i].ds_addr, 419 dma_seg[i].ds_len, 420 eop, 421 desc); 422 if (eop) 423 break; 424 i++; 425 sfxge_next_stmp(txq, &stmp); 426 } 427 txq->n_pend_desc = n_dma_seg + vlan_tagged; 428 } 429 430 /* 431 * If the mapping required more than one descriptor 432 * then we need to associate the DMA map with the last 433 * descriptor, not the first. 434 */ 435 if (used_map != &stmp->map) { 436 map = stmp->map; 437 stmp->map = *used_map; 438 *used_map = map; 439 } 440 441 stmp->u.mbuf = mbuf; 442 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; 443 444 /* Post the fragment list. */ 445 sfxge_tx_qlist_post(txq); 446 447 return (0); 448 449 reject_mapped: 450 bus_dmamap_unload(txq->packet_dma_tag, *used_map); 451 reject: 452 /* Drop the packet on the floor. */ 453 m_freem(mbuf); 454 ++txq->drops; 455 456 return (rc); 457 } 458 459 /* 460 * Drain the deferred packet list into the transmit queue. 461 */ 462 static void 463 sfxge_tx_qdpl_drain(struct sfxge_txq *txq) 464 { 465 struct sfxge_softc *sc; 466 struct sfxge_tx_dpl *stdp; 467 struct mbuf *mbuf, *next; 468 unsigned int count; 469 unsigned int non_tcp_count; 470 unsigned int pushed; 471 int rc; 472 473 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 474 475 sc = txq->sc; 476 stdp = &txq->dpl; 477 pushed = txq->added; 478 479 if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) { 480 prefetch_read_many(sc->enp); 481 prefetch_read_many(txq->common); 482 } 483 484 mbuf = stdp->std_get; 485 count = stdp->std_get_count; 486 non_tcp_count = stdp->std_get_non_tcp_count; 487 488 if (count > stdp->std_get_hiwat) 489 stdp->std_get_hiwat = count; 490 491 while (count != 0) { 492 KASSERT(mbuf != NULL, ("mbuf == NULL")); 493 494 next = mbuf->m_nextpkt; 495 mbuf->m_nextpkt = NULL; 496 497 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ 498 499 if (next != NULL) 500 prefetch_read_many(next); 501 502 rc = sfxge_tx_queue_mbuf(txq, mbuf); 503 --count; 504 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); 505 mbuf = next; 506 if (rc != 0) 507 continue; 508 509 if (txq->blocked) 510 break; 511 512 /* Push the fragments to the hardware in batches. */ 513 if (txq->added - pushed >= SFXGE_TX_BATCH) { 514 efx_tx_qpush(txq->common, txq->added, pushed); 515 pushed = txq->added; 516 } 517 } 518 519 if (count == 0) { 520 KASSERT(mbuf == NULL, ("mbuf != NULL")); 521 KASSERT(non_tcp_count == 0, 522 ("inconsistent TCP/non-TCP detection")); 523 stdp->std_get = NULL; 524 stdp->std_get_count = 0; 525 stdp->std_get_non_tcp_count = 0; 526 stdp->std_getp = &stdp->std_get; 527 } else { 528 stdp->std_get = mbuf; 529 stdp->std_get_count = count; 530 stdp->std_get_non_tcp_count = non_tcp_count; 531 } 532 533 if (txq->added != pushed) 534 efx_tx_qpush(txq->common, txq->added, pushed); 535 536 KASSERT(txq->blocked || stdp->std_get_count == 0, 537 ("queue unblocked but count is non-zero")); 538 } 539 540 #define SFXGE_TX_QDPL_PENDING(_txq) ((_txq)->dpl.std_put != 0) 541 542 /* 543 * Service the deferred packet list. 544 * 545 * NOTE: drops the txq mutex! 546 */ 547 static void 548 sfxge_tx_qdpl_service(struct sfxge_txq *txq) 549 { 550 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 551 552 do { 553 if (SFXGE_TX_QDPL_PENDING(txq)) 554 sfxge_tx_qdpl_swizzle(txq); 555 556 if (!txq->blocked) 557 sfxge_tx_qdpl_drain(txq); 558 559 SFXGE_TXQ_UNLOCK(txq); 560 } while (SFXGE_TX_QDPL_PENDING(txq) && 561 SFXGE_TXQ_TRYLOCK(txq)); 562 } 563 564 /* 565 * Put a packet on the deferred packet get-list. 566 */ 567 static int 568 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf) 569 { 570 struct sfxge_tx_dpl *stdp; 571 572 stdp = &txq->dpl; 573 574 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 575 576 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 577 578 if (stdp->std_get_count >= stdp->std_get_max) { 579 txq->get_overflow++; 580 return (ENOBUFS); 581 } 582 if (sfxge_is_mbuf_non_tcp(mbuf)) { 583 if (stdp->std_get_non_tcp_count >= 584 stdp->std_get_non_tcp_max) { 585 txq->get_non_tcp_overflow++; 586 return (ENOBUFS); 587 } 588 stdp->std_get_non_tcp_count++; 589 } 590 591 *(stdp->std_getp) = mbuf; 592 stdp->std_getp = &mbuf->m_nextpkt; 593 stdp->std_get_count++; 594 595 return (0); 596 } 597 598 /* 599 * Put a packet on the deferred packet put-list. 600 * 601 * We overload the csum_data field in the mbuf to keep track of this length 602 * because there is no cheap alternative to avoid races. 603 */ 604 static int 605 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf) 606 { 607 struct sfxge_tx_dpl *stdp; 608 volatile uintptr_t *putp; 609 uintptr_t old; 610 uintptr_t new; 611 unsigned old_len; 612 613 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 614 615 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 616 617 stdp = &txq->dpl; 618 putp = &stdp->std_put; 619 new = (uintptr_t)mbuf; 620 621 do { 622 old = *putp; 623 if (old != 0) { 624 struct mbuf *mp = (struct mbuf *)old; 625 old_len = mp->m_pkthdr.csum_data; 626 } else 627 old_len = 0; 628 if (old_len >= stdp->std_put_max) { 629 atomic_add_long(&txq->put_overflow, 1); 630 return (ENOBUFS); 631 } 632 mbuf->m_pkthdr.csum_data = old_len + 1; 633 mbuf->m_nextpkt = (void *)old; 634 } while (atomic_cmpset_ptr(putp, old, new) == 0); 635 636 return (0); 637 } 638 639 /* 640 * Called from if_transmit - will try to grab the txq lock and enqueue to the 641 * put list if it succeeds, otherwise try to push onto the defer list if space. 642 */ 643 static int 644 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) 645 { 646 int rc; 647 648 if (!SFXGE_LINK_UP(txq->sc)) { 649 atomic_add_long(&txq->netdown_drops, 1); 650 return (ENETDOWN); 651 } 652 653 /* 654 * Try to grab the txq lock. If we are able to get the lock, 655 * the packet will be appended to the "get list" of the deferred 656 * packet list. Otherwise, it will be pushed on the "put list". 657 */ 658 if (SFXGE_TXQ_TRYLOCK(txq)) { 659 /* First swizzle put-list to get-list to keep order */ 660 sfxge_tx_qdpl_swizzle(txq); 661 662 rc = sfxge_tx_qdpl_put_locked(txq, m); 663 664 /* Try to service the list. */ 665 sfxge_tx_qdpl_service(txq); 666 /* Lock has been dropped. */ 667 } else { 668 rc = sfxge_tx_qdpl_put_unlocked(txq, m); 669 670 /* 671 * Try to grab the lock again. 672 * 673 * If we are able to get the lock, we need to process 674 * the deferred packet list. If we are not able to get 675 * the lock, another thread is processing the list. 676 */ 677 if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) { 678 sfxge_tx_qdpl_service(txq); 679 /* Lock has been dropped. */ 680 } 681 } 682 683 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 684 685 return (rc); 686 } 687 688 static void 689 sfxge_tx_qdpl_flush(struct sfxge_txq *txq) 690 { 691 struct sfxge_tx_dpl *stdp = &txq->dpl; 692 struct mbuf *mbuf, *next; 693 694 SFXGE_TXQ_LOCK(txq); 695 696 sfxge_tx_qdpl_swizzle(txq); 697 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { 698 next = mbuf->m_nextpkt; 699 m_freem(mbuf); 700 } 701 stdp->std_get = NULL; 702 stdp->std_get_count = 0; 703 stdp->std_get_non_tcp_count = 0; 704 stdp->std_getp = &stdp->std_get; 705 706 SFXGE_TXQ_UNLOCK(txq); 707 } 708 709 void 710 sfxge_if_qflush(struct ifnet *ifp) 711 { 712 struct sfxge_softc *sc; 713 unsigned int i; 714 715 sc = ifp->if_softc; 716 717 for (i = 0; i < sc->txq_count; i++) 718 sfxge_tx_qdpl_flush(sc->txq[i]); 719 } 720 721 #if SFXGE_TX_PARSE_EARLY 722 723 /* There is little space for user data in mbuf pkthdr, so we 724 * use l*hlen fields which are not used by the driver otherwise 725 * to store header offsets. 726 * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes. 727 */ 728 729 730 #define TSO_MBUF_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[0]) 731 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */ 732 #define TSO_MBUF_FLAGS(_mbuf) ((_mbuf)->m_pkthdr.l5hlen) 733 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1]) 734 #define TSO_MBUF_SEQNUM(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1]) 735 736 static void sfxge_parse_tx_packet(struct mbuf *mbuf) 737 { 738 struct ether_header *eh = mtod(mbuf, struct ether_header *); 739 const struct tcphdr *th; 740 struct tcphdr th_copy; 741 742 /* Find network protocol and header */ 743 TSO_MBUF_PROTO(mbuf) = eh->ether_type; 744 if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) { 745 struct ether_vlan_header *veh = 746 mtod(mbuf, struct ether_vlan_header *); 747 TSO_MBUF_PROTO(mbuf) = veh->evl_proto; 748 mbuf->m_pkthdr.l2hlen = sizeof(*veh); 749 } else { 750 mbuf->m_pkthdr.l2hlen = sizeof(*eh); 751 } 752 753 /* Find TCP header */ 754 if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) { 755 const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen); 756 757 KASSERT(iph->ip_p == IPPROTO_TCP, 758 ("TSO required on non-TCP packet")); 759 mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl; 760 TSO_MBUF_PACKETID(mbuf) = iph->ip_id; 761 } else { 762 KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6), 763 ("TSO required on non-IP packet")); 764 KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt == 765 IPPROTO_TCP, 766 ("TSO required on non-TCP packet")); 767 mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr); 768 TSO_MBUF_PACKETID(mbuf) = 0; 769 } 770 771 KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen, 772 ("network header is fragmented in mbuf")); 773 774 /* We need TCP header including flags (window is the next) */ 775 if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) { 776 m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy), 777 (caddr_t)&th_copy); 778 th = &th_copy; 779 } else { 780 th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen); 781 } 782 783 mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off; 784 TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq); 785 786 /* These flags must not be duplicated */ 787 /* 788 * RST should not be duplicated as well, but FreeBSD kernel 789 * generates TSO packets with RST flag. So, do not assert 790 * its absence. 791 */ 792 KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), 793 ("incompatible TCP flag 0x%x on TSO packet", 794 th->th_flags & (TH_URG | TH_SYN))); 795 TSO_MBUF_FLAGS(mbuf) = th->th_flags; 796 } 797 #endif 798 799 /* 800 * TX start -- called by the stack. 801 */ 802 int 803 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) 804 { 805 struct sfxge_softc *sc; 806 struct sfxge_txq *txq; 807 int rc; 808 809 sc = (struct sfxge_softc *)ifp->if_softc; 810 811 /* 812 * Transmit may be called when interface is up from the kernel 813 * point of view, but not yet up (in progress) from the driver 814 * point of view. I.e. link aggregation bring up. 815 * Transmit may be called when interface is up from the driver 816 * point of view, but already down from the kernel point of 817 * view. I.e. Rx when interface shutdown is in progress. 818 */ 819 KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP), 820 ("interface not up")); 821 822 /* Pick the desired transmit queue. */ 823 if (m->m_pkthdr.csum_flags & 824 (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) { 825 int index = 0; 826 827 #ifdef RSS 828 uint32_t bucket_id; 829 830 /* 831 * Select a TX queue which matches the corresponding 832 * RX queue for the hash in order to assign both 833 * TX and RX parts of the flow to the same CPU 834 */ 835 if (rss_m2bucket(m, &bucket_id) == 0) 836 index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1)); 837 #else 838 /* check if flowid is set */ 839 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 840 uint32_t hash = m->m_pkthdr.flowid; 841 842 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX]; 843 } 844 #endif 845 #if SFXGE_TX_PARSE_EARLY 846 if (m->m_pkthdr.csum_flags & CSUM_TSO) 847 sfxge_parse_tx_packet(m); 848 #endif 849 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; 850 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { 851 txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; 852 } else { 853 txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; 854 } 855 856 rc = sfxge_tx_packet_add(txq, m); 857 if (rc != 0) 858 m_freem(m); 859 860 return (rc); 861 } 862 863 /* 864 * Software "TSO". Not quite as good as doing it in hardware, but 865 * still faster than segmenting in the stack. 866 */ 867 868 struct sfxge_tso_state { 869 /* Output position */ 870 unsigned out_len; /* Remaining length in current segment */ 871 unsigned seqnum; /* Current sequence number */ 872 unsigned packet_space; /* Remaining space in current packet */ 873 unsigned segs_space; /* Remaining number of DMA segments 874 for the packet (FATSOv2 only) */ 875 876 /* Input position */ 877 uint64_t dma_addr; /* DMA address of current position */ 878 unsigned in_len; /* Remaining length in current mbuf */ 879 880 const struct mbuf *mbuf; /* Input mbuf (head of chain) */ 881 u_short protocol; /* Network protocol (after VLAN decap) */ 882 ssize_t nh_off; /* Offset of network header */ 883 ssize_t tcph_off; /* Offset of TCP header */ 884 unsigned header_len; /* Number of bytes of header */ 885 unsigned seg_size; /* TCP segment size */ 886 int fw_assisted; /* Use FW-assisted TSO */ 887 u_short packet_id; /* IPv4 packet ID from the original packet */ 888 uint8_t tcp_flags; /* TCP flags */ 889 efx_desc_t header_desc; /* Precomputed header descriptor for 890 * FW-assisted TSO */ 891 }; 892 893 #if !SFXGE_TX_PARSE_EARLY 894 static const struct ip *tso_iph(const struct sfxge_tso_state *tso) 895 { 896 KASSERT(tso->protocol == htons(ETHERTYPE_IP), 897 ("tso_iph() in non-IPv4 state")); 898 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); 899 } 900 901 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) 902 { 903 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 904 ("tso_ip6h() in non-IPv6 state")); 905 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); 906 } 907 908 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) 909 { 910 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); 911 } 912 #endif 913 914 915 /* Size of preallocated TSO header buffers. Larger blocks must be 916 * allocated from the heap. 917 */ 918 #define TSOH_STD_SIZE 128 919 920 /* At most half the descriptors in the queue at any time will refer to 921 * a TSO header buffer, since they must always be followed by a 922 * payload descriptor referring to an mbuf. 923 */ 924 #define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) 925 #define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 926 #define TSOH_PAGE_COUNT(_txq_entries) \ 927 howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE) 928 929 static int tso_init(struct sfxge_txq *txq) 930 { 931 struct sfxge_softc *sc = txq->sc; 932 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); 933 int i, rc; 934 935 /* Allocate TSO header buffers */ 936 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), 937 M_SFXGE, M_WAITOK); 938 939 for (i = 0; i < tsoh_page_count; i++) { 940 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); 941 if (rc != 0) 942 goto fail; 943 } 944 945 return (0); 946 947 fail: 948 while (i-- > 0) 949 sfxge_dma_free(&txq->tsoh_buffer[i]); 950 free(txq->tsoh_buffer, M_SFXGE); 951 txq->tsoh_buffer = NULL; 952 return (rc); 953 } 954 955 static void tso_fini(struct sfxge_txq *txq) 956 { 957 int i; 958 959 if (txq->tsoh_buffer != NULL) { 960 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) 961 sfxge_dma_free(&txq->tsoh_buffer[i]); 962 free(txq->tsoh_buffer, M_SFXGE); 963 } 964 } 965 966 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso, 967 const bus_dma_segment_t *hdr_dma_seg, 968 struct mbuf *mbuf) 969 { 970 const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp); 971 #if !SFXGE_TX_PARSE_EARLY 972 struct ether_header *eh = mtod(mbuf, struct ether_header *); 973 const struct tcphdr *th; 974 struct tcphdr th_copy; 975 #endif 976 977 tso->fw_assisted = txq->tso_fw_assisted; 978 tso->mbuf = mbuf; 979 980 /* Find network protocol and header */ 981 #if !SFXGE_TX_PARSE_EARLY 982 tso->protocol = eh->ether_type; 983 if (tso->protocol == htons(ETHERTYPE_VLAN)) { 984 struct ether_vlan_header *veh = 985 mtod(mbuf, struct ether_vlan_header *); 986 tso->protocol = veh->evl_proto; 987 tso->nh_off = sizeof(*veh); 988 } else { 989 tso->nh_off = sizeof(*eh); 990 } 991 #else 992 tso->protocol = TSO_MBUF_PROTO(mbuf); 993 tso->nh_off = mbuf->m_pkthdr.l2hlen; 994 tso->tcph_off = mbuf->m_pkthdr.l3hlen; 995 tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf)); 996 #endif 997 998 #if !SFXGE_TX_PARSE_EARLY 999 /* Find TCP header */ 1000 if (tso->protocol == htons(ETHERTYPE_IP)) { 1001 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, 1002 ("TSO required on non-TCP packet")); 1003 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; 1004 tso->packet_id = ntohs(tso_iph(tso)->ip_id); 1005 } else { 1006 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 1007 ("TSO required on non-IP packet")); 1008 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, 1009 ("TSO required on non-TCP packet")); 1010 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); 1011 tso->packet_id = 0; 1012 } 1013 #endif 1014 1015 1016 if (tso->fw_assisted && 1017 __predict_false(tso->tcph_off > 1018 encp->enc_tx_tso_tcp_header_offset_limit)) { 1019 tso->fw_assisted = 0; 1020 } 1021 1022 1023 #if !SFXGE_TX_PARSE_EARLY 1024 KASSERT(mbuf->m_len >= tso->tcph_off, 1025 ("network header is fragmented in mbuf")); 1026 /* We need TCP header including flags (window is the next) */ 1027 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) { 1028 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy), 1029 (caddr_t)&th_copy); 1030 th = &th_copy; 1031 } else { 1032 th = tso_tcph(tso); 1033 } 1034 tso->header_len = tso->tcph_off + 4 * th->th_off; 1035 #else 1036 tso->header_len = mbuf->m_pkthdr.l4hlen; 1037 #endif 1038 tso->seg_size = mbuf->m_pkthdr.tso_segsz; 1039 1040 #if !SFXGE_TX_PARSE_EARLY 1041 tso->seqnum = ntohl(th->th_seq); 1042 1043 /* These flags must not be duplicated */ 1044 /* 1045 * RST should not be duplicated as well, but FreeBSD kernel 1046 * generates TSO packets with RST flag. So, do not assert 1047 * its absence. 1048 */ 1049 KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), 1050 ("incompatible TCP flag 0x%x on TSO packet", 1051 th->th_flags & (TH_URG | TH_SYN))); 1052 tso->tcp_flags = th->th_flags; 1053 #else 1054 tso->seqnum = TSO_MBUF_SEQNUM(mbuf); 1055 tso->tcp_flags = TSO_MBUF_FLAGS(mbuf); 1056 #endif 1057 1058 tso->out_len = mbuf->m_pkthdr.len - tso->header_len; 1059 1060 if (tso->fw_assisted) { 1061 if (hdr_dma_seg->ds_len >= tso->header_len) 1062 efx_tx_qdesc_dma_create(txq->common, 1063 hdr_dma_seg->ds_addr, 1064 tso->header_len, 1065 B_FALSE, 1066 &tso->header_desc); 1067 else 1068 tso->fw_assisted = 0; 1069 } 1070 } 1071 1072 /* 1073 * tso_fill_packet_with_fragment - form descriptors for the current fragment 1074 * 1075 * Form descriptors for the current fragment, until we reach the end 1076 * of fragment or end-of-packet. Return 0 on success, 1 if not enough 1077 * space. 1078 */ 1079 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, 1080 struct sfxge_tso_state *tso) 1081 { 1082 efx_desc_t *desc; 1083 int n; 1084 uint64_t dma_addr = tso->dma_addr; 1085 boolean_t eop; 1086 1087 if (tso->in_len == 0 || tso->packet_space == 0) 1088 return; 1089 1090 KASSERT(tso->in_len > 0, ("TSO input length went negative")); 1091 KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); 1092 1093 if (tso->fw_assisted & SFXGE_FATSOV2) { 1094 n = tso->in_len; 1095 tso->out_len -= n; 1096 tso->seqnum += n; 1097 tso->in_len = 0; 1098 if (n < tso->packet_space) { 1099 tso->packet_space -= n; 1100 tso->segs_space--; 1101 } else { 1102 tso->packet_space = tso->seg_size - 1103 (n - tso->packet_space) % tso->seg_size; 1104 tso->segs_space = 1105 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 - 1106 (tso->packet_space != tso->seg_size); 1107 } 1108 } else { 1109 n = min(tso->in_len, tso->packet_space); 1110 tso->packet_space -= n; 1111 tso->out_len -= n; 1112 tso->dma_addr += n; 1113 tso->in_len -= n; 1114 } 1115 1116 /* 1117 * It is OK to use binary OR below to avoid extra branching 1118 * since all conditions may always be checked. 1119 */ 1120 eop = (tso->out_len == 0) | (tso->packet_space == 0) | 1121 (tso->segs_space == 0); 1122 1123 desc = &txq->pend_desc[txq->n_pend_desc++]; 1124 efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc); 1125 } 1126 1127 /* Callback from bus_dmamap_load() for long TSO headers. */ 1128 static void tso_map_long_header(void *dma_addr_ret, 1129 bus_dma_segment_t *segs, int nseg, 1130 int error) 1131 { 1132 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && 1133 __predict_true(nseg == 1)) ? 1134 segs->ds_addr : 0); 1135 } 1136 1137 /* 1138 * tso_start_new_packet - generate a new header and prepare for the new packet 1139 * 1140 * Generate a new header and prepare for the new packet. Return 0 on 1141 * success, or an error code if failed to alloc header. 1142 */ 1143 static int tso_start_new_packet(struct sfxge_txq *txq, 1144 struct sfxge_tso_state *tso, 1145 unsigned int *idp) 1146 { 1147 unsigned int id = *idp; 1148 struct tcphdr *tsoh_th; 1149 unsigned ip_length; 1150 caddr_t header; 1151 uint64_t dma_addr; 1152 bus_dmamap_t map; 1153 efx_desc_t *desc; 1154 int rc; 1155 1156 if (tso->fw_assisted) { 1157 if (tso->fw_assisted & SFXGE_FATSOV2) { 1158 /* Add 2 FATSOv2 option descriptors */ 1159 desc = &txq->pend_desc[txq->n_pend_desc]; 1160 efx_tx_qdesc_tso2_create(txq->common, 1161 tso->packet_id, 1162 tso->seqnum, 1163 tso->seg_size, 1164 desc, 1165 EFX_TX_FATSOV2_OPT_NDESCS); 1166 desc += EFX_TX_FATSOV2_OPT_NDESCS; 1167 txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS; 1168 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1169 id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask; 1170 1171 tso->segs_space = 1172 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1; 1173 } else { 1174 uint8_t tcp_flags = tso->tcp_flags; 1175 1176 if (tso->out_len > tso->seg_size) 1177 tcp_flags &= ~(TH_FIN | TH_PUSH); 1178 1179 /* Add FATSOv1 option descriptor */ 1180 desc = &txq->pend_desc[txq->n_pend_desc++]; 1181 efx_tx_qdesc_tso_create(txq->common, 1182 tso->packet_id, 1183 tso->seqnum, 1184 tcp_flags, 1185 desc++); 1186 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1187 id = (id + 1) & txq->ptr_mask; 1188 1189 tso->seqnum += tso->seg_size; 1190 tso->segs_space = UINT_MAX; 1191 } 1192 1193 /* Header DMA descriptor */ 1194 *desc = tso->header_desc; 1195 txq->n_pend_desc++; 1196 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1197 id = (id + 1) & txq->ptr_mask; 1198 } else { 1199 /* Allocate a DMA-mapped header buffer. */ 1200 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { 1201 unsigned int page_index = (id / 2) / TSOH_PER_PAGE; 1202 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; 1203 1204 header = (txq->tsoh_buffer[page_index].esm_base + 1205 buf_index * TSOH_STD_SIZE); 1206 dma_addr = (txq->tsoh_buffer[page_index].esm_addr + 1207 buf_index * TSOH_STD_SIZE); 1208 map = txq->tsoh_buffer[page_index].esm_map; 1209 1210 KASSERT(txq->stmp[id].flags == 0, 1211 ("stmp flags are not 0")); 1212 } else { 1213 struct sfxge_tx_mapping *stmp = &txq->stmp[id]; 1214 1215 /* We cannot use bus_dmamem_alloc() as that may sleep */ 1216 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); 1217 if (__predict_false(!header)) 1218 return (ENOMEM); 1219 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, 1220 header, tso->header_len, 1221 tso_map_long_header, &dma_addr, 1222 BUS_DMA_NOWAIT); 1223 if (__predict_false(dma_addr == 0)) { 1224 if (rc == 0) { 1225 /* Succeeded but got >1 segment */ 1226 bus_dmamap_unload(txq->packet_dma_tag, 1227 stmp->map); 1228 rc = EINVAL; 1229 } 1230 free(header, M_SFXGE); 1231 return (rc); 1232 } 1233 map = stmp->map; 1234 1235 txq->tso_long_headers++; 1236 stmp->u.heap_buf = header; 1237 stmp->flags = TX_BUF_UNMAP; 1238 } 1239 1240 tsoh_th = (struct tcphdr *)(header + tso->tcph_off); 1241 1242 /* Copy and update the headers. */ 1243 m_copydata(tso->mbuf, 0, tso->header_len, header); 1244 1245 tsoh_th->th_seq = htonl(tso->seqnum); 1246 tso->seqnum += tso->seg_size; 1247 if (tso->out_len > tso->seg_size) { 1248 /* This packet will not finish the TSO burst. */ 1249 ip_length = tso->header_len - tso->nh_off + tso->seg_size; 1250 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); 1251 } else { 1252 /* This packet will be the last in the TSO burst. */ 1253 ip_length = tso->header_len - tso->nh_off + tso->out_len; 1254 } 1255 1256 if (tso->protocol == htons(ETHERTYPE_IP)) { 1257 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); 1258 tsoh_iph->ip_len = htons(ip_length); 1259 /* XXX We should increment ip_id, but FreeBSD doesn't 1260 * currently allocate extra IDs for multiple segments. 1261 */ 1262 } else { 1263 struct ip6_hdr *tsoh_iph = 1264 (struct ip6_hdr *)(header + tso->nh_off); 1265 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); 1266 } 1267 1268 /* Make the header visible to the hardware. */ 1269 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); 1270 1271 /* Form a descriptor for this header. */ 1272 desc = &txq->pend_desc[txq->n_pend_desc++]; 1273 efx_tx_qdesc_dma_create(txq->common, 1274 dma_addr, 1275 tso->header_len, 1276 0, 1277 desc); 1278 id = (id + 1) & txq->ptr_mask; 1279 1280 tso->segs_space = UINT_MAX; 1281 } 1282 tso->packet_space = tso->seg_size; 1283 txq->tso_packets++; 1284 *idp = id; 1285 1286 return (0); 1287 } 1288 1289 static int 1290 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 1291 const bus_dma_segment_t *dma_seg, int n_dma_seg, 1292 int vlan_tagged) 1293 { 1294 struct sfxge_tso_state tso; 1295 unsigned int id; 1296 unsigned skipped = 0; 1297 1298 tso_start(txq, &tso, dma_seg, mbuf); 1299 1300 while (dma_seg->ds_len + skipped <= tso.header_len) { 1301 skipped += dma_seg->ds_len; 1302 --n_dma_seg; 1303 KASSERT(n_dma_seg, ("no payload found in TSO packet")); 1304 ++dma_seg; 1305 } 1306 tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); 1307 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); 1308 1309 id = (txq->added + vlan_tagged) & txq->ptr_mask; 1310 if (__predict_false(tso_start_new_packet(txq, &tso, &id))) 1311 return (-1); 1312 1313 while (1) { 1314 tso_fill_packet_with_fragment(txq, &tso); 1315 /* Exactly one DMA descriptor is added */ 1316 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1317 id = (id + 1) & txq->ptr_mask; 1318 1319 /* Move onto the next fragment? */ 1320 if (tso.in_len == 0) { 1321 --n_dma_seg; 1322 if (n_dma_seg == 0) 1323 break; 1324 ++dma_seg; 1325 tso.in_len = dma_seg->ds_len; 1326 tso.dma_addr = dma_seg->ds_addr; 1327 } 1328 1329 /* End of packet? */ 1330 if ((tso.packet_space == 0) | (tso.segs_space == 0)) { 1331 unsigned int n_fatso_opt_desc = 1332 (tso.fw_assisted & SFXGE_FATSOV2) ? 1333 EFX_TX_FATSOV2_OPT_NDESCS : 1334 (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0; 1335 1336 /* If the queue is now full due to tiny MSS, 1337 * or we can't create another header, discard 1338 * the remainder of the input mbuf but do not 1339 * roll back the work we have done. 1340 */ 1341 if (txq->n_pend_desc + n_fatso_opt_desc + 1342 1 /* header */ + n_dma_seg > txq->max_pkt_desc) { 1343 txq->tso_pdrop_too_many++; 1344 break; 1345 } 1346 if (__predict_false(tso_start_new_packet(txq, &tso, 1347 &id))) { 1348 txq->tso_pdrop_no_rsrc++; 1349 break; 1350 } 1351 } 1352 } 1353 1354 txq->tso_bursts++; 1355 return (id); 1356 } 1357 1358 static void 1359 sfxge_tx_qunblock(struct sfxge_txq *txq) 1360 { 1361 struct sfxge_softc *sc; 1362 struct sfxge_evq *evq; 1363 1364 sc = txq->sc; 1365 evq = sc->evq[txq->evq_index]; 1366 1367 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 1368 1369 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) 1370 return; 1371 1372 SFXGE_TXQ_LOCK(txq); 1373 1374 if (txq->blocked) { 1375 unsigned int level; 1376 1377 level = txq->added - txq->completed; 1378 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) { 1379 /* reaped must be in sync with blocked */ 1380 sfxge_tx_qreap(txq); 1381 txq->blocked = 0; 1382 } 1383 } 1384 1385 sfxge_tx_qdpl_service(txq); 1386 /* note: lock has been dropped */ 1387 } 1388 1389 void 1390 sfxge_tx_qflush_done(struct sfxge_txq *txq) 1391 { 1392 1393 txq->flush_state = SFXGE_FLUSH_DONE; 1394 } 1395 1396 static void 1397 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) 1398 { 1399 struct sfxge_txq *txq; 1400 struct sfxge_evq *evq; 1401 unsigned int count; 1402 1403 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1404 1405 txq = sc->txq[index]; 1406 evq = sc->evq[txq->evq_index]; 1407 1408 SFXGE_EVQ_LOCK(evq); 1409 SFXGE_TXQ_LOCK(txq); 1410 1411 KASSERT(txq->init_state == SFXGE_TXQ_STARTED, 1412 ("txq->init_state != SFXGE_TXQ_STARTED")); 1413 1414 txq->init_state = SFXGE_TXQ_INITIALIZED; 1415 1416 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1417 txq->flush_state = SFXGE_FLUSH_PENDING; 1418 1419 SFXGE_EVQ_UNLOCK(evq); 1420 SFXGE_TXQ_UNLOCK(txq); 1421 1422 /* Flush the transmit queue. */ 1423 if (efx_tx_qflush(txq->common) != 0) { 1424 log(LOG_ERR, "%s: Flushing Tx queue %u failed\n", 1425 device_get_nameunit(sc->dev), index); 1426 txq->flush_state = SFXGE_FLUSH_DONE; 1427 } else { 1428 count = 0; 1429 do { 1430 /* Spin for 100ms. */ 1431 DELAY(100000); 1432 if (txq->flush_state != SFXGE_FLUSH_PENDING) 1433 break; 1434 } while (++count < 20); 1435 } 1436 SFXGE_EVQ_LOCK(evq); 1437 SFXGE_TXQ_LOCK(txq); 1438 1439 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, 1440 ("txq->flush_state == SFXGE_FLUSH_FAILED")); 1441 1442 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1443 /* Flush timeout */ 1444 log(LOG_ERR, "%s: Cannot flush Tx queue %u\n", 1445 device_get_nameunit(sc->dev), index); 1446 txq->flush_state = SFXGE_FLUSH_DONE; 1447 } 1448 } 1449 1450 txq->blocked = 0; 1451 txq->pending = txq->added; 1452 1453 sfxge_tx_qcomplete(txq, evq); 1454 KASSERT(txq->completed == txq->added, 1455 ("txq->completed != txq->added")); 1456 1457 sfxge_tx_qreap(txq); 1458 KASSERT(txq->reaped == txq->completed, 1459 ("txq->reaped != txq->completed")); 1460 1461 txq->added = 0; 1462 txq->pending = 0; 1463 txq->completed = 0; 1464 txq->reaped = 0; 1465 1466 /* Destroy the common code transmit queue. */ 1467 efx_tx_qdestroy(txq->common); 1468 txq->common = NULL; 1469 1470 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1471 EFX_TXQ_NBUFS(sc->txq_entries)); 1472 1473 SFXGE_EVQ_UNLOCK(evq); 1474 SFXGE_TXQ_UNLOCK(txq); 1475 } 1476 1477 /* 1478 * Estimate maximum number of Tx descriptors required for TSO packet. 1479 * With minimum MSS and maximum mbuf length we might need more (even 1480 * than a ring-ful of descriptors), but this should not happen in 1481 * practice except due to deliberate attack. In that case we will 1482 * truncate the output at a packet boundary. 1483 */ 1484 static unsigned int 1485 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type, 1486 unsigned int tso_fw_assisted) 1487 { 1488 /* One descriptor for every input fragment */ 1489 unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG; 1490 unsigned int sw_tso_max_descs; 1491 unsigned int fa_tso_v1_max_descs = 0; 1492 unsigned int fa_tso_v2_max_descs = 0; 1493 1494 /* VLAN tagging Tx option descriptor may be required */ 1495 if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled) 1496 max_descs++; 1497 1498 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) { 1499 /* 1500 * Plus header and payload descriptor for each output segment. 1501 * Minus one since header fragment is already counted. 1502 * Even if FATSO is used, we should be ready to fallback 1503 * to do it in the driver. 1504 */ 1505 sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1; 1506 1507 /* FW assisted TSOv1 requires one more descriptor per segment 1508 * in comparison to SW TSO */ 1509 if (tso_fw_assisted & SFXGE_FATSOV1) 1510 fa_tso_v1_max_descs = 1511 sw_tso_max_descs + SFXGE_TSO_MAX_SEGS; 1512 1513 /* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra 1514 * descriptors per superframe limited by number of DMA fetches 1515 * per packet. The first packet header is already counted. 1516 */ 1517 if (tso_fw_assisted & SFXGE_FATSOV2) { 1518 fa_tso_v2_max_descs = 1519 howmany(SFXGE_TX_MAPPING_MAX_SEG, 1520 EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) * 1521 (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1; 1522 } 1523 1524 max_descs += MAX(sw_tso_max_descs, 1525 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs)); 1526 } 1527 1528 return (max_descs); 1529 } 1530 1531 static int 1532 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) 1533 { 1534 struct sfxge_txq *txq; 1535 efsys_mem_t *esmp; 1536 uint16_t flags; 1537 unsigned int tso_fw_assisted; 1538 struct sfxge_evq *evq; 1539 unsigned int desc_index; 1540 int rc; 1541 1542 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1543 1544 txq = sc->txq[index]; 1545 esmp = &txq->mem; 1546 evq = sc->evq[txq->evq_index]; 1547 1548 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1549 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1550 KASSERT(evq->init_state == SFXGE_EVQ_STARTED, 1551 ("evq->init_state != SFXGE_EVQ_STARTED")); 1552 1553 /* Program the buffer table. */ 1554 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, 1555 EFX_TXQ_NBUFS(sc->txq_entries))) != 0) 1556 return (rc); 1557 1558 /* Determine the kind of queue we are creating. */ 1559 tso_fw_assisted = 0; 1560 switch (txq->type) { 1561 case SFXGE_TXQ_NON_CKSUM: 1562 flags = 0; 1563 break; 1564 case SFXGE_TXQ_IP_CKSUM: 1565 flags = EFX_TXQ_CKSUM_IPV4; 1566 break; 1567 case SFXGE_TXQ_IP_TCP_UDP_CKSUM: 1568 flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP; 1569 tso_fw_assisted = sc->tso_fw_assisted; 1570 if (tso_fw_assisted & SFXGE_FATSOV2) 1571 flags |= EFX_TXQ_FATSOV2; 1572 break; 1573 default: 1574 KASSERT(0, ("Impossible TX queue")); 1575 flags = 0; 1576 break; 1577 } 1578 1579 /* Create the common code transmit queue. */ 1580 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1581 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1582 &txq->common, &desc_index)) != 0) { 1583 /* Retry if no FATSOv2 resources, otherwise fail */ 1584 if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2)) 1585 goto fail; 1586 1587 /* Looks like all FATSOv2 contexts are used */ 1588 flags &= ~EFX_TXQ_FATSOV2; 1589 tso_fw_assisted &= ~SFXGE_FATSOV2; 1590 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1591 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1592 &txq->common, &desc_index)) != 0) 1593 goto fail; 1594 } 1595 1596 /* Initialise queue descriptor indexes */ 1597 txq->added = txq->pending = txq->completed = txq->reaped = desc_index; 1598 1599 SFXGE_TXQ_LOCK(txq); 1600 1601 /* Enable the transmit queue. */ 1602 efx_tx_qenable(txq->common); 1603 1604 txq->init_state = SFXGE_TXQ_STARTED; 1605 txq->flush_state = SFXGE_FLUSH_REQUIRED; 1606 txq->tso_fw_assisted = tso_fw_assisted; 1607 1608 txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type, 1609 tso_fw_assisted); 1610 1611 SFXGE_TXQ_UNLOCK(txq); 1612 1613 return (0); 1614 1615 fail: 1616 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1617 EFX_TXQ_NBUFS(sc->txq_entries)); 1618 return (rc); 1619 } 1620 1621 void 1622 sfxge_tx_stop(struct sfxge_softc *sc) 1623 { 1624 int index; 1625 1626 index = sc->txq_count; 1627 while (--index >= 0) 1628 sfxge_tx_qstop(sc, index); 1629 1630 /* Tear down the transmit module */ 1631 efx_tx_fini(sc->enp); 1632 } 1633 1634 int 1635 sfxge_tx_start(struct sfxge_softc *sc) 1636 { 1637 int index; 1638 int rc; 1639 1640 /* Initialize the common code transmit module. */ 1641 if ((rc = efx_tx_init(sc->enp)) != 0) 1642 return (rc); 1643 1644 for (index = 0; index < sc->txq_count; index++) { 1645 if ((rc = sfxge_tx_qstart(sc, index)) != 0) 1646 goto fail; 1647 } 1648 1649 return (0); 1650 1651 fail: 1652 while (--index >= 0) 1653 sfxge_tx_qstop(sc, index); 1654 1655 efx_tx_fini(sc->enp); 1656 1657 return (rc); 1658 } 1659 1660 static int 1661 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node) 1662 { 1663 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev); 1664 struct sysctl_oid *stat_node; 1665 unsigned int id; 1666 1667 stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1668 "stats", CTLFLAG_RD, NULL, 1669 "Tx queue statistics"); 1670 if (stat_node == NULL) 1671 return (ENOMEM); 1672 1673 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1674 SYSCTL_ADD_ULONG( 1675 ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO, 1676 sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS, 1677 (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset), 1678 ""); 1679 } 1680 1681 return (0); 1682 } 1683 1684 /** 1685 * Destroy a transmit queue. 1686 */ 1687 static void 1688 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) 1689 { 1690 struct sfxge_txq *txq; 1691 unsigned int nmaps; 1692 1693 txq = sc->txq[index]; 1694 1695 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1696 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1697 1698 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) 1699 tso_fini(txq); 1700 1701 /* Free the context arrays. */ 1702 free(txq->pend_desc, M_SFXGE); 1703 nmaps = sc->txq_entries; 1704 while (nmaps-- != 0) 1705 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1706 free(txq->stmp, M_SFXGE); 1707 1708 /* Release DMA memory mapping. */ 1709 sfxge_dma_free(&txq->mem); 1710 1711 sc->txq[index] = NULL; 1712 1713 SFXGE_TXQ_LOCK_DESTROY(txq); 1714 1715 free(txq, M_SFXGE); 1716 } 1717 1718 static int 1719 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, 1720 enum sfxge_txq_type type, unsigned int evq_index) 1721 { 1722 char name[16]; 1723 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1724 struct sysctl_oid *txq_node; 1725 struct sfxge_txq *txq; 1726 struct sfxge_evq *evq; 1727 struct sfxge_tx_dpl *stdp; 1728 struct sysctl_oid *dpl_node; 1729 efsys_mem_t *esmp; 1730 unsigned int nmaps; 1731 int rc; 1732 1733 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); 1734 txq->sc = sc; 1735 txq->entries = sc->txq_entries; 1736 txq->ptr_mask = txq->entries - 1; 1737 1738 sc->txq[txq_index] = txq; 1739 esmp = &txq->mem; 1740 1741 evq = sc->evq[evq_index]; 1742 1743 /* Allocate and zero DMA space for the descriptor ring. */ 1744 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) 1745 return (rc); 1746 1747 /* Allocate buffer table entries. */ 1748 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), 1749 &txq->buf_base_id); 1750 1751 /* Create a DMA tag for packet mappings. */ 1752 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000, 1753 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, 1754 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL, 1755 &txq->packet_dma_tag) != 0) { 1756 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); 1757 rc = ENOMEM; 1758 goto fail; 1759 } 1760 1761 /* Allocate pending descriptor array for batching writes. */ 1762 txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries, 1763 M_SFXGE, M_ZERO | M_WAITOK); 1764 1765 /* Allocate and initialise mbuf DMA mapping array. */ 1766 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, 1767 M_SFXGE, M_ZERO | M_WAITOK); 1768 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { 1769 rc = bus_dmamap_create(txq->packet_dma_tag, 0, 1770 &txq->stmp[nmaps].map); 1771 if (rc != 0) 1772 goto fail2; 1773 } 1774 1775 snprintf(name, sizeof(name), "%u", txq_index); 1776 txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node), 1777 OID_AUTO, name, CTLFLAG_RD, NULL, ""); 1778 if (txq_node == NULL) { 1779 rc = ENOMEM; 1780 goto fail_txq_node; 1781 } 1782 1783 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && 1784 (rc = tso_init(txq)) != 0) 1785 goto fail3; 1786 1787 if (sfxge_tx_dpl_get_max <= 0) { 1788 log(LOG_ERR, "%s=%d must be greater than 0", 1789 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); 1790 rc = EINVAL; 1791 goto fail_tx_dpl_get_max; 1792 } 1793 if (sfxge_tx_dpl_get_non_tcp_max <= 0) { 1794 log(LOG_ERR, "%s=%d must be greater than 0", 1795 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, 1796 sfxge_tx_dpl_get_non_tcp_max); 1797 rc = EINVAL; 1798 goto fail_tx_dpl_get_max; 1799 } 1800 if (sfxge_tx_dpl_put_max < 0) { 1801 log(LOG_ERR, "%s=%d must be greater or equal to 0", 1802 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); 1803 rc = EINVAL; 1804 goto fail_tx_dpl_put_max; 1805 } 1806 1807 /* Initialize the deferred packet list. */ 1808 stdp = &txq->dpl; 1809 stdp->std_put_max = sfxge_tx_dpl_put_max; 1810 stdp->std_get_max = sfxge_tx_dpl_get_max; 1811 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; 1812 stdp->std_getp = &stdp->std_get; 1813 1814 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index); 1815 1816 dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1817 "dpl", CTLFLAG_RD, NULL, 1818 "Deferred packet list statistics"); 1819 if (dpl_node == NULL) { 1820 rc = ENOMEM; 1821 goto fail_dpl_node; 1822 } 1823 1824 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1825 "get_count", CTLFLAG_RD | CTLFLAG_STATS, 1826 &stdp->std_get_count, 0, ""); 1827 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1828 "get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, 1829 &stdp->std_get_non_tcp_count, 0, ""); 1830 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1831 "get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1832 &stdp->std_get_hiwat, 0, ""); 1833 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1834 "put_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1835 &stdp->std_put_hiwat, 0, ""); 1836 1837 rc = sfxge_txq_stat_init(txq, txq_node); 1838 if (rc != 0) 1839 goto fail_txq_stat_init; 1840 1841 txq->type = type; 1842 txq->evq_index = evq_index; 1843 txq->txq_index = txq_index; 1844 txq->init_state = SFXGE_TXQ_INITIALIZED; 1845 txq->hw_vlan_tci = 0; 1846 1847 return (0); 1848 1849 fail_txq_stat_init: 1850 fail_dpl_node: 1851 fail_tx_dpl_put_max: 1852 fail_tx_dpl_get_max: 1853 fail3: 1854 fail_txq_node: 1855 free(txq->pend_desc, M_SFXGE); 1856 fail2: 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 1862 fail: 1863 sfxge_dma_free(esmp); 1864 1865 return (rc); 1866 } 1867 1868 static int 1869 sfxge_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 1885 static void 1886 sfxge_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 1904 uint64_t 1905 sfxge_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 1927 void 1928 sfxge_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 1940 int 1941 sfxge_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 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc; 1954 1955 sc->tso_fw_assisted = sfxge_tso_fw_assisted; 1956 if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) || 1957 (!encp->enc_fw_assisted_tso_enabled)) 1958 sc->tso_fw_assisted &= ~SFXGE_FATSOV1; 1959 if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) || 1960 (!encp->enc_fw_assisted_tso_v2_enabled)) 1961 sc->tso_fw_assisted &= ~SFXGE_FATSOV2; 1962 1963 sc->txqs_node = SYSCTL_ADD_NODE( 1964 device_get_sysctl_ctx(sc->dev), 1965 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), 1966 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); 1967 if (sc->txqs_node == NULL) { 1968 rc = ENOMEM; 1969 goto fail_txq_node; 1970 } 1971 1972 /* Initialize the transmit queues */ 1973 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, 1974 SFXGE_TXQ_NON_CKSUM, 0)) != 0) 1975 goto fail; 1976 1977 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, 1978 SFXGE_TXQ_IP_CKSUM, 0)) != 0) 1979 goto fail2; 1980 1981 for (index = 0; 1982 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1; 1983 index++) { 1984 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index, 1985 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) 1986 goto fail3; 1987 } 1988 1989 sfxge_tx_stat_init(sc); 1990 1991 return (0); 1992 1993 fail3: 1994 while (--index >= 0) 1995 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1996 1997 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 1998 1999 fail2: 2000 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 2001 2002 fail: 2003 fail_txq_node: 2004 sc->txq_count = 0; 2005 return (rc); 2006 } 2007