1 /*- 2 * Copyright (c) 2010-2015 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 <sys/types.h> 53 #include <sys/mbuf.h> 54 #include <sys/smp.h> 55 #include <sys/socket.h> 56 #include <sys/sysctl.h> 57 #include <sys/syslog.h> 58 59 #include <net/bpf.h> 60 #include <net/ethernet.h> 61 #include <net/if.h> 62 #include <net/if_vlan_var.h> 63 64 #include <netinet/in.h> 65 #include <netinet/ip.h> 66 #include <netinet/ip6.h> 67 #include <netinet/tcp.h> 68 69 #include "common/efx.h" 70 71 #include "sfxge.h" 72 #include "sfxge_tx.h" 73 74 75 #define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max) 76 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT; 77 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max); 78 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN, 79 &sfxge_tx_dpl_get_max, 0, 80 "Maximum number of any packets in deferred packet get-list"); 81 82 #define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \ 83 SFXGE_PARAM(tx_dpl_get_non_tcp_max) 84 static int sfxge_tx_dpl_get_non_tcp_max = 85 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT; 86 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max); 87 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN, 88 &sfxge_tx_dpl_get_non_tcp_max, 0, 89 "Maximum number of non-TCP packets in deferred packet get-list"); 90 91 #define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max) 92 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT; 93 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max); 94 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN, 95 &sfxge_tx_dpl_put_max, 0, 96 "Maximum number of any packets in deferred packet put-list"); 97 98 #define SFXGE_PARAM_TSO_FW_ASSISTED SFXGE_PARAM(tso_fw_assisted) 99 static int sfxge_tso_fw_assisted = 1; 100 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted); 101 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN, 102 &sfxge_tso_fw_assisted, 0, 103 "Use FW-assisted TSO if supported by NIC firmware"); 104 105 106 static const struct { 107 const char *name; 108 size_t offset; 109 } sfxge_tx_stats[] = { 110 #define SFXGE_TX_STAT(name, member) \ 111 { #name, offsetof(struct sfxge_txq, member) } 112 SFXGE_TX_STAT(tso_bursts, tso_bursts), 113 SFXGE_TX_STAT(tso_packets, tso_packets), 114 SFXGE_TX_STAT(tso_long_headers, tso_long_headers), 115 SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many), 116 SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc), 117 SFXGE_TX_STAT(tx_collapses, collapses), 118 SFXGE_TX_STAT(tx_drops, drops), 119 SFXGE_TX_STAT(tx_get_overflow, get_overflow), 120 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow), 121 SFXGE_TX_STAT(tx_put_overflow, put_overflow), 122 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops), 123 }; 124 125 126 /* Forward declarations. */ 127 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq); 128 static void sfxge_tx_qlist_post(struct sfxge_txq *txq); 129 static void sfxge_tx_qunblock(struct sfxge_txq *txq); 130 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 131 const bus_dma_segment_t *dma_seg, int n_dma_seg, 132 int vlan_tagged); 133 134 static int 135 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf) 136 { 137 uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ? 138 mbuf->m_pkthdr.ether_vtag : 139 0); 140 141 if (this_tag == txq->hw_vlan_tci) 142 return (0); 143 144 efx_tx_qdesc_vlantci_create(txq->common, 145 bswap16(this_tag), 146 &txq->pend_desc[0]); 147 txq->n_pend_desc = 1; 148 txq->hw_vlan_tci = this_tag; 149 return (1); 150 } 151 152 static inline void 153 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp) 154 { 155 KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0")); 156 if (__predict_false(*pstmp == 157 &txq->stmp[txq->ptr_mask])) 158 *pstmp = &txq->stmp[0]; 159 else 160 (*pstmp)++; 161 } 162 163 164 void 165 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) 166 { 167 unsigned int completed; 168 169 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 170 171 completed = txq->completed; 172 while (completed != txq->pending) { 173 struct sfxge_tx_mapping *stmp; 174 unsigned int id; 175 176 id = completed++ & txq->ptr_mask; 177 178 stmp = &txq->stmp[id]; 179 if (stmp->flags & TX_BUF_UNMAP) { 180 bus_dmamap_unload(txq->packet_dma_tag, stmp->map); 181 if (stmp->flags & TX_BUF_MBUF) { 182 struct mbuf *m = stmp->u.mbuf; 183 do 184 m = m_free(m); 185 while (m != NULL); 186 } else { 187 free(stmp->u.heap_buf, M_SFXGE); 188 } 189 stmp->flags = 0; 190 } 191 } 192 txq->completed = completed; 193 194 /* Check whether we need to unblock the queue. */ 195 mb(); 196 if (txq->blocked) { 197 unsigned int level; 198 199 level = txq->added - txq->completed; 200 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) 201 sfxge_tx_qunblock(txq); 202 } 203 } 204 205 static unsigned int 206 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf) 207 { 208 /* Absense of TCP checksum flags does not mean that it is non-TCP 209 * but it should be true if user wants to achieve high throughput. 210 */ 211 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP))); 212 } 213 214 /* 215 * Reorder the put list and append it to the get list. 216 */ 217 static void 218 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq) 219 { 220 struct sfxge_tx_dpl *stdp; 221 struct mbuf *mbuf, *get_next, **get_tailp; 222 volatile uintptr_t *putp; 223 uintptr_t put; 224 unsigned int count; 225 unsigned int non_tcp_count; 226 227 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 228 229 stdp = &txq->dpl; 230 231 /* Acquire the put list. */ 232 putp = &stdp->std_put; 233 put = atomic_readandclear_ptr(putp); 234 mbuf = (void *)put; 235 236 if (mbuf == NULL) 237 return; 238 239 /* Reverse the put list. */ 240 get_tailp = &mbuf->m_nextpkt; 241 get_next = NULL; 242 243 count = 0; 244 non_tcp_count = 0; 245 do { 246 struct mbuf *put_next; 247 248 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf); 249 put_next = mbuf->m_nextpkt; 250 mbuf->m_nextpkt = get_next; 251 get_next = mbuf; 252 mbuf = put_next; 253 254 count++; 255 } while (mbuf != NULL); 256 257 if (count > stdp->std_put_hiwat) 258 stdp->std_put_hiwat = count; 259 260 /* Append the reversed put list to the get list. */ 261 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL")); 262 *stdp->std_getp = get_next; 263 stdp->std_getp = get_tailp; 264 stdp->std_get_count += count; 265 stdp->std_get_non_tcp_count += non_tcp_count; 266 } 267 268 static void 269 sfxge_tx_qreap(struct sfxge_txq *txq) 270 { 271 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 272 273 txq->reaped = txq->completed; 274 } 275 276 static void 277 sfxge_tx_qlist_post(struct sfxge_txq *txq) 278 { 279 unsigned int old_added; 280 unsigned int block_level; 281 unsigned int level; 282 int rc; 283 284 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 285 286 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0")); 287 KASSERT(txq->n_pend_desc <= txq->max_pkt_desc, 288 ("txq->n_pend_desc too large")); 289 KASSERT(!txq->blocked, ("txq->blocked")); 290 291 old_added = txq->added; 292 293 /* Post the fragment list. */ 294 rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc, 295 txq->reaped, &txq->added); 296 KASSERT(rc == 0, ("efx_tx_qdesc_post() failed")); 297 298 /* If efx_tx_qdesc_post() had to refragment, our information about 299 * buffers to free may be associated with the wrong 300 * descriptors. 301 */ 302 KASSERT(txq->added - old_added == txq->n_pend_desc, 303 ("efx_tx_qdesc_post() refragmented descriptors")); 304 305 level = txq->added - txq->reaped; 306 KASSERT(level <= txq->entries, ("overfilled TX queue")); 307 308 /* Clear the fragment list. */ 309 txq->n_pend_desc = 0; 310 311 /* 312 * Set the block level to ensure there is space to generate a 313 * large number of descriptors for TSO. 314 */ 315 block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc; 316 317 /* Have we reached the block level? */ 318 if (level < block_level) 319 return; 320 321 /* Reap, and check again */ 322 sfxge_tx_qreap(txq); 323 level = txq->added - txq->reaped; 324 if (level < block_level) 325 return; 326 327 txq->blocked = 1; 328 329 /* 330 * Avoid a race with completion interrupt handling that could leave 331 * the queue blocked. 332 */ 333 mb(); 334 sfxge_tx_qreap(txq); 335 level = txq->added - txq->reaped; 336 if (level < block_level) { 337 mb(); 338 txq->blocked = 0; 339 } 340 } 341 342 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf) 343 { 344 bus_dmamap_t *used_map; 345 bus_dmamap_t map; 346 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG]; 347 unsigned int id; 348 struct sfxge_tx_mapping *stmp; 349 efx_desc_t *desc; 350 int n_dma_seg; 351 int rc; 352 int i; 353 int eop; 354 int vlan_tagged; 355 356 KASSERT(!txq->blocked, ("txq->blocked")); 357 358 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) 359 prefetch_read_many(mbuf->m_data); 360 361 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) { 362 rc = EINTR; 363 goto reject; 364 } 365 366 /* Load the packet for DMA. */ 367 id = txq->added & txq->ptr_mask; 368 stmp = &txq->stmp[id]; 369 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, 370 mbuf, dma_seg, &n_dma_seg, 0); 371 if (rc == EFBIG) { 372 /* Try again. */ 373 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, 374 SFXGE_TX_MAPPING_MAX_SEG); 375 if (new_mbuf == NULL) 376 goto reject; 377 ++txq->collapses; 378 mbuf = new_mbuf; 379 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, 380 stmp->map, mbuf, 381 dma_seg, &n_dma_seg, 0); 382 } 383 if (rc != 0) 384 goto reject; 385 386 /* Make the packet visible to the hardware. */ 387 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); 388 389 used_map = &stmp->map; 390 391 vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf); 392 if (vlan_tagged) { 393 sfxge_next_stmp(txq, &stmp); 394 } 395 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 396 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged); 397 if (rc < 0) 398 goto reject_mapped; 399 stmp = &txq->stmp[(rc - 1) & txq->ptr_mask]; 400 } else { 401 /* Add the mapping to the fragment list, and set flags 402 * for the buffer. 403 */ 404 405 i = 0; 406 for (;;) { 407 desc = &txq->pend_desc[i + vlan_tagged]; 408 eop = (i == n_dma_seg - 1); 409 efx_tx_qdesc_dma_create(txq->common, 410 dma_seg[i].ds_addr, 411 dma_seg[i].ds_len, 412 eop, 413 desc); 414 if (eop) 415 break; 416 i++; 417 sfxge_next_stmp(txq, &stmp); 418 } 419 txq->n_pend_desc = n_dma_seg + vlan_tagged; 420 } 421 422 /* 423 * If the mapping required more than one descriptor 424 * then we need to associate the DMA map with the last 425 * descriptor, not the first. 426 */ 427 if (used_map != &stmp->map) { 428 map = stmp->map; 429 stmp->map = *used_map; 430 *used_map = map; 431 } 432 433 stmp->u.mbuf = mbuf; 434 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; 435 436 /* Post the fragment list. */ 437 sfxge_tx_qlist_post(txq); 438 439 return (0); 440 441 reject_mapped: 442 bus_dmamap_unload(txq->packet_dma_tag, *used_map); 443 reject: 444 /* Drop the packet on the floor. */ 445 m_freem(mbuf); 446 ++txq->drops; 447 448 return (rc); 449 } 450 451 /* 452 * Drain the deferred packet list into the transmit queue. 453 */ 454 static void 455 sfxge_tx_qdpl_drain(struct sfxge_txq *txq) 456 { 457 struct sfxge_softc *sc; 458 struct sfxge_tx_dpl *stdp; 459 struct mbuf *mbuf, *next; 460 unsigned int count; 461 unsigned int non_tcp_count; 462 unsigned int pushed; 463 int rc; 464 465 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 466 467 sc = txq->sc; 468 stdp = &txq->dpl; 469 pushed = txq->added; 470 471 if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) { 472 prefetch_read_many(sc->enp); 473 prefetch_read_many(txq->common); 474 } 475 476 mbuf = stdp->std_get; 477 count = stdp->std_get_count; 478 non_tcp_count = stdp->std_get_non_tcp_count; 479 480 if (count > stdp->std_get_hiwat) 481 stdp->std_get_hiwat = count; 482 483 while (count != 0) { 484 KASSERT(mbuf != NULL, ("mbuf == NULL")); 485 486 next = mbuf->m_nextpkt; 487 mbuf->m_nextpkt = NULL; 488 489 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ 490 491 if (next != NULL) 492 prefetch_read_many(next); 493 494 rc = sfxge_tx_queue_mbuf(txq, mbuf); 495 --count; 496 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); 497 mbuf = next; 498 if (rc != 0) 499 continue; 500 501 if (txq->blocked) 502 break; 503 504 /* Push the fragments to the hardware in batches. */ 505 if (txq->added - pushed >= SFXGE_TX_BATCH) { 506 efx_tx_qpush(txq->common, txq->added, pushed); 507 pushed = txq->added; 508 } 509 } 510 511 if (count == 0) { 512 KASSERT(mbuf == NULL, ("mbuf != NULL")); 513 KASSERT(non_tcp_count == 0, 514 ("inconsistent TCP/non-TCP detection")); 515 stdp->std_get = NULL; 516 stdp->std_get_count = 0; 517 stdp->std_get_non_tcp_count = 0; 518 stdp->std_getp = &stdp->std_get; 519 } else { 520 stdp->std_get = mbuf; 521 stdp->std_get_count = count; 522 stdp->std_get_non_tcp_count = non_tcp_count; 523 } 524 525 if (txq->added != pushed) 526 efx_tx_qpush(txq->common, txq->added, pushed); 527 528 KASSERT(txq->blocked || stdp->std_get_count == 0, 529 ("queue unblocked but count is non-zero")); 530 } 531 532 #define SFXGE_TX_QDPL_PENDING(_txq) ((_txq)->dpl.std_put != 0) 533 534 /* 535 * Service the deferred packet list. 536 * 537 * NOTE: drops the txq mutex! 538 */ 539 static void 540 sfxge_tx_qdpl_service(struct sfxge_txq *txq) 541 { 542 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 543 544 do { 545 if (SFXGE_TX_QDPL_PENDING(txq)) 546 sfxge_tx_qdpl_swizzle(txq); 547 548 if (!txq->blocked) 549 sfxge_tx_qdpl_drain(txq); 550 551 SFXGE_TXQ_UNLOCK(txq); 552 } while (SFXGE_TX_QDPL_PENDING(txq) && 553 SFXGE_TXQ_TRYLOCK(txq)); 554 } 555 556 /* 557 * Put a packet on the deferred packet get-list. 558 */ 559 static int 560 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf) 561 { 562 struct sfxge_tx_dpl *stdp; 563 564 stdp = &txq->dpl; 565 566 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 567 568 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 569 570 if (stdp->std_get_count >= stdp->std_get_max) { 571 txq->get_overflow++; 572 return (ENOBUFS); 573 } 574 if (sfxge_is_mbuf_non_tcp(mbuf)) { 575 if (stdp->std_get_non_tcp_count >= 576 stdp->std_get_non_tcp_max) { 577 txq->get_non_tcp_overflow++; 578 return (ENOBUFS); 579 } 580 stdp->std_get_non_tcp_count++; 581 } 582 583 *(stdp->std_getp) = mbuf; 584 stdp->std_getp = &mbuf->m_nextpkt; 585 stdp->std_get_count++; 586 587 return (0); 588 } 589 590 /* 591 * Put a packet on the deferred packet put-list. 592 * 593 * We overload the csum_data field in the mbuf to keep track of this length 594 * because there is no cheap alternative to avoid races. 595 */ 596 static int 597 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf) 598 { 599 struct sfxge_tx_dpl *stdp; 600 volatile uintptr_t *putp; 601 uintptr_t old; 602 uintptr_t new; 603 unsigned old_len; 604 605 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 606 607 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 608 609 stdp = &txq->dpl; 610 putp = &stdp->std_put; 611 new = (uintptr_t)mbuf; 612 613 do { 614 old = *putp; 615 if (old != 0) { 616 struct mbuf *mp = (struct mbuf *)old; 617 old_len = mp->m_pkthdr.csum_data; 618 } else 619 old_len = 0; 620 if (old_len >= stdp->std_put_max) { 621 atomic_add_long(&txq->put_overflow, 1); 622 return (ENOBUFS); 623 } 624 mbuf->m_pkthdr.csum_data = old_len + 1; 625 mbuf->m_nextpkt = (void *)old; 626 } while (atomic_cmpset_ptr(putp, old, new) == 0); 627 628 return (0); 629 } 630 631 /* 632 * Called from if_transmit - will try to grab the txq lock and enqueue to the 633 * put list if it succeeds, otherwise try to push onto the defer list if space. 634 */ 635 static int 636 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) 637 { 638 int rc; 639 640 if (!SFXGE_LINK_UP(txq->sc)) { 641 atomic_add_long(&txq->netdown_drops, 1); 642 return (ENETDOWN); 643 } 644 645 /* 646 * Try to grab the txq lock. If we are able to get the lock, 647 * the packet will be appended to the "get list" of the deferred 648 * packet list. Otherwise, it will be pushed on the "put list". 649 */ 650 if (SFXGE_TXQ_TRYLOCK(txq)) { 651 /* First swizzle put-list to get-list to keep order */ 652 sfxge_tx_qdpl_swizzle(txq); 653 654 rc = sfxge_tx_qdpl_put_locked(txq, m); 655 656 /* Try to service the list. */ 657 sfxge_tx_qdpl_service(txq); 658 /* Lock has been dropped. */ 659 } else { 660 rc = sfxge_tx_qdpl_put_unlocked(txq, m); 661 662 /* 663 * Try to grab the lock again. 664 * 665 * If we are able to get the lock, we need to process 666 * the deferred packet list. If we are not able to get 667 * the lock, another thread is processing the list. 668 */ 669 if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) { 670 sfxge_tx_qdpl_service(txq); 671 /* Lock has been dropped. */ 672 } 673 } 674 675 SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); 676 677 return (rc); 678 } 679 680 static void 681 sfxge_tx_qdpl_flush(struct sfxge_txq *txq) 682 { 683 struct sfxge_tx_dpl *stdp = &txq->dpl; 684 struct mbuf *mbuf, *next; 685 686 SFXGE_TXQ_LOCK(txq); 687 688 sfxge_tx_qdpl_swizzle(txq); 689 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { 690 next = mbuf->m_nextpkt; 691 m_freem(mbuf); 692 } 693 stdp->std_get = NULL; 694 stdp->std_get_count = 0; 695 stdp->std_get_non_tcp_count = 0; 696 stdp->std_getp = &stdp->std_get; 697 698 SFXGE_TXQ_UNLOCK(txq); 699 } 700 701 void 702 sfxge_if_qflush(struct ifnet *ifp) 703 { 704 struct sfxge_softc *sc; 705 unsigned int i; 706 707 sc = ifp->if_softc; 708 709 for (i = 0; i < sc->txq_count; i++) 710 sfxge_tx_qdpl_flush(sc->txq[i]); 711 } 712 713 /* 714 * TX start -- called by the stack. 715 */ 716 int 717 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) 718 { 719 struct sfxge_softc *sc; 720 struct sfxge_txq *txq; 721 int rc; 722 723 sc = (struct sfxge_softc *)ifp->if_softc; 724 725 /* 726 * Transmit may be called when interface is up from the kernel 727 * point of view, but not yet up (in progress) from the driver 728 * point of view. I.e. link aggregation bring up. 729 * Transmit may be called when interface is up from the driver 730 * point of view, but already down from the kernel point of 731 * view. I.e. Rx when interface shutdown is in progress. 732 */ 733 KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP), 734 ("interface not up")); 735 736 /* Pick the desired transmit queue. */ 737 if (m->m_pkthdr.csum_flags & 738 (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) { 739 int index = 0; 740 741 /* check if flowid is set */ 742 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 743 uint32_t hash = m->m_pkthdr.flowid; 744 745 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX]; 746 } 747 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; 748 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { 749 txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; 750 } else { 751 txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; 752 } 753 754 rc = sfxge_tx_packet_add(txq, m); 755 if (rc != 0) 756 m_freem(m); 757 758 return (rc); 759 } 760 761 /* 762 * Software "TSO". Not quite as good as doing it in hardware, but 763 * still faster than segmenting in the stack. 764 */ 765 766 struct sfxge_tso_state { 767 /* Output position */ 768 unsigned out_len; /* Remaining length in current segment */ 769 unsigned seqnum; /* Current sequence number */ 770 unsigned packet_space; /* Remaining space in current packet */ 771 772 /* Input position */ 773 uint64_t dma_addr; /* DMA address of current position */ 774 unsigned in_len; /* Remaining length in current mbuf */ 775 776 const struct mbuf *mbuf; /* Input mbuf (head of chain) */ 777 u_short protocol; /* Network protocol (after VLAN decap) */ 778 ssize_t nh_off; /* Offset of network header */ 779 ssize_t tcph_off; /* Offset of TCP header */ 780 unsigned header_len; /* Number of bytes of header */ 781 unsigned seg_size; /* TCP segment size */ 782 int fw_assisted; /* Use FW-assisted TSO */ 783 u_short packet_id; /* IPv4 packet ID from the original packet */ 784 efx_desc_t header_desc; /* Precomputed header descriptor for 785 * FW-assisted TSO */ 786 }; 787 788 static const struct ip *tso_iph(const struct sfxge_tso_state *tso) 789 { 790 KASSERT(tso->protocol == htons(ETHERTYPE_IP), 791 ("tso_iph() in non-IPv4 state")); 792 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); 793 } 794 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) 795 { 796 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 797 ("tso_ip6h() in non-IPv6 state")); 798 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); 799 } 800 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) 801 { 802 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); 803 } 804 805 /* Size of preallocated TSO header buffers. Larger blocks must be 806 * allocated from the heap. 807 */ 808 #define TSOH_STD_SIZE 128 809 810 /* At most half the descriptors in the queue at any time will refer to 811 * a TSO header buffer, since they must always be followed by a 812 * payload descriptor referring to an mbuf. 813 */ 814 #define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) 815 #define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 816 #define TSOH_PAGE_COUNT(_txq_entries) \ 817 ((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE) 818 819 static int tso_init(struct sfxge_txq *txq) 820 { 821 struct sfxge_softc *sc = txq->sc; 822 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); 823 int i, rc; 824 825 /* Allocate TSO header buffers */ 826 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), 827 M_SFXGE, M_WAITOK); 828 829 for (i = 0; i < tsoh_page_count; i++) { 830 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); 831 if (rc != 0) 832 goto fail; 833 } 834 835 return (0); 836 837 fail: 838 while (i-- > 0) 839 sfxge_dma_free(&txq->tsoh_buffer[i]); 840 free(txq->tsoh_buffer, M_SFXGE); 841 txq->tsoh_buffer = NULL; 842 return (rc); 843 } 844 845 static void tso_fini(struct sfxge_txq *txq) 846 { 847 int i; 848 849 if (txq->tsoh_buffer != NULL) { 850 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) 851 sfxge_dma_free(&txq->tsoh_buffer[i]); 852 free(txq->tsoh_buffer, M_SFXGE); 853 } 854 } 855 856 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso, 857 const bus_dma_segment_t *hdr_dma_seg, 858 struct mbuf *mbuf) 859 { 860 struct ether_header *eh = mtod(mbuf, struct ether_header *); 861 const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp); 862 const struct tcphdr *th; 863 struct tcphdr th_copy; 864 865 tso->fw_assisted = txq->sc->tso_fw_assisted; 866 tso->mbuf = mbuf; 867 868 /* Find network protocol and header */ 869 tso->protocol = eh->ether_type; 870 if (tso->protocol == htons(ETHERTYPE_VLAN)) { 871 struct ether_vlan_header *veh = 872 mtod(mbuf, struct ether_vlan_header *); 873 tso->protocol = veh->evl_proto; 874 tso->nh_off = sizeof(*veh); 875 } else { 876 tso->nh_off = sizeof(*eh); 877 } 878 879 /* Find TCP header */ 880 if (tso->protocol == htons(ETHERTYPE_IP)) { 881 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, 882 ("TSO required on non-TCP packet")); 883 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; 884 tso->packet_id = tso_iph(tso)->ip_id; 885 } else { 886 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 887 ("TSO required on non-IP packet")); 888 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, 889 ("TSO required on non-TCP packet")); 890 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); 891 tso->packet_id = 0; 892 } 893 if (tso->fw_assisted && 894 __predict_false(tso->tcph_off > 895 encp->enc_tx_tso_tcp_header_offset_limit)) { 896 tso->fw_assisted = 0; 897 } 898 899 KASSERT(mbuf->m_len >= tso->tcph_off, 900 ("network header is fragmented in mbuf")); 901 /* We need TCP header including flags (window is the next) */ 902 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) { 903 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy), 904 (caddr_t)&th_copy); 905 th = &th_copy; 906 } else { 907 th = tso_tcph(tso); 908 } 909 910 tso->header_len = tso->tcph_off + 4 * th->th_off; 911 tso->seg_size = mbuf->m_pkthdr.tso_segsz; 912 913 tso->seqnum = ntohl(th->th_seq); 914 915 /* These flags must not be duplicated */ 916 /* 917 * RST should not be duplicated as well, but FreeBSD kernel 918 * generates TSO packets with RST flag. So, do not assert 919 * its absence. 920 */ 921 KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), 922 ("incompatible TCP flag 0x%x on TSO packet", 923 th->th_flags & (TH_URG | TH_SYN))); 924 925 tso->out_len = mbuf->m_pkthdr.len - tso->header_len; 926 927 if (tso->fw_assisted) { 928 if (hdr_dma_seg->ds_len >= tso->header_len) 929 efx_tx_qdesc_dma_create(txq->common, 930 hdr_dma_seg->ds_addr, 931 tso->header_len, 932 B_FALSE, 933 &tso->header_desc); 934 else 935 tso->fw_assisted = 0; 936 } 937 } 938 939 /* 940 * tso_fill_packet_with_fragment - form descriptors for the current fragment 941 * 942 * Form descriptors for the current fragment, until we reach the end 943 * of fragment or end-of-packet. Return 0 on success, 1 if not enough 944 * space. 945 */ 946 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, 947 struct sfxge_tso_state *tso) 948 { 949 efx_desc_t *desc; 950 int n; 951 952 if (tso->in_len == 0 || tso->packet_space == 0) 953 return; 954 955 KASSERT(tso->in_len > 0, ("TSO input length went negative")); 956 KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); 957 958 n = min(tso->in_len, tso->packet_space); 959 960 tso->packet_space -= n; 961 tso->out_len -= n; 962 tso->in_len -= n; 963 964 desc = &txq->pend_desc[txq->n_pend_desc++]; 965 efx_tx_qdesc_dma_create(txq->common, 966 tso->dma_addr, 967 n, 968 tso->out_len == 0 || tso->packet_space == 0, 969 desc); 970 971 tso->dma_addr += n; 972 } 973 974 /* Callback from bus_dmamap_load() for long TSO headers. */ 975 static void tso_map_long_header(void *dma_addr_ret, 976 bus_dma_segment_t *segs, int nseg, 977 int error) 978 { 979 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && 980 __predict_true(nseg == 1)) ? 981 segs->ds_addr : 0); 982 } 983 984 /* 985 * tso_start_new_packet - generate a new header and prepare for the new packet 986 * 987 * Generate a new header and prepare for the new packet. Return 0 on 988 * success, or an error code if failed to alloc header. 989 */ 990 static int tso_start_new_packet(struct sfxge_txq *txq, 991 struct sfxge_tso_state *tso, 992 unsigned int *idp) 993 { 994 unsigned int id = *idp; 995 struct tcphdr *tsoh_th; 996 unsigned ip_length; 997 caddr_t header; 998 uint64_t dma_addr; 999 bus_dmamap_t map; 1000 efx_desc_t *desc; 1001 int rc; 1002 1003 if (tso->fw_assisted) { 1004 uint8_t tcp_flags = tso_tcph(tso)->th_flags; 1005 1006 if (tso->out_len > tso->seg_size) 1007 tcp_flags &= ~(TH_FIN | TH_PUSH); 1008 1009 /* TSO option descriptor */ 1010 desc = &txq->pend_desc[txq->n_pend_desc++]; 1011 efx_tx_qdesc_tso_create(txq->common, 1012 tso->packet_id, 1013 tso->seqnum, 1014 tcp_flags, 1015 desc++); 1016 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1017 id = (id + 1) & txq->ptr_mask; 1018 1019 /* Header DMA descriptor */ 1020 *desc = tso->header_desc; 1021 txq->n_pend_desc++; 1022 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1023 id = (id + 1) & txq->ptr_mask; 1024 1025 tso->seqnum += tso->seg_size; 1026 } else { 1027 /* Allocate a DMA-mapped header buffer. */ 1028 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { 1029 unsigned int page_index = (id / 2) / TSOH_PER_PAGE; 1030 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; 1031 1032 header = (txq->tsoh_buffer[page_index].esm_base + 1033 buf_index * TSOH_STD_SIZE); 1034 dma_addr = (txq->tsoh_buffer[page_index].esm_addr + 1035 buf_index * TSOH_STD_SIZE); 1036 map = txq->tsoh_buffer[page_index].esm_map; 1037 1038 KASSERT(txq->stmp[id].flags == 0, 1039 ("stmp flags are not 0")); 1040 } else { 1041 struct sfxge_tx_mapping *stmp = &txq->stmp[id]; 1042 1043 /* We cannot use bus_dmamem_alloc() as that may sleep */ 1044 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); 1045 if (__predict_false(!header)) 1046 return (ENOMEM); 1047 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, 1048 header, tso->header_len, 1049 tso_map_long_header, &dma_addr, 1050 BUS_DMA_NOWAIT); 1051 if (__predict_false(dma_addr == 0)) { 1052 if (rc == 0) { 1053 /* Succeeded but got >1 segment */ 1054 bus_dmamap_unload(txq->packet_dma_tag, 1055 stmp->map); 1056 rc = EINVAL; 1057 } 1058 free(header, M_SFXGE); 1059 return (rc); 1060 } 1061 map = stmp->map; 1062 1063 txq->tso_long_headers++; 1064 stmp->u.heap_buf = header; 1065 stmp->flags = TX_BUF_UNMAP; 1066 } 1067 1068 tsoh_th = (struct tcphdr *)(header + tso->tcph_off); 1069 1070 /* Copy and update the headers. */ 1071 m_copydata(tso->mbuf, 0, tso->header_len, header); 1072 1073 tsoh_th->th_seq = htonl(tso->seqnum); 1074 tso->seqnum += tso->seg_size; 1075 if (tso->out_len > tso->seg_size) { 1076 /* This packet will not finish the TSO burst. */ 1077 ip_length = tso->header_len - tso->nh_off + tso->seg_size; 1078 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); 1079 } else { 1080 /* This packet will be the last in the TSO burst. */ 1081 ip_length = tso->header_len - tso->nh_off + tso->out_len; 1082 } 1083 1084 if (tso->protocol == htons(ETHERTYPE_IP)) { 1085 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); 1086 tsoh_iph->ip_len = htons(ip_length); 1087 /* XXX We should increment ip_id, but FreeBSD doesn't 1088 * currently allocate extra IDs for multiple segments. 1089 */ 1090 } else { 1091 struct ip6_hdr *tsoh_iph = 1092 (struct ip6_hdr *)(header + tso->nh_off); 1093 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); 1094 } 1095 1096 /* Make the header visible to the hardware. */ 1097 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); 1098 1099 /* Form a descriptor for this header. */ 1100 desc = &txq->pend_desc[txq->n_pend_desc++]; 1101 efx_tx_qdesc_dma_create(txq->common, 1102 dma_addr, 1103 tso->header_len, 1104 0, 1105 desc); 1106 id = (id + 1) & txq->ptr_mask; 1107 } 1108 tso->packet_space = tso->seg_size; 1109 txq->tso_packets++; 1110 *idp = id; 1111 1112 return (0); 1113 } 1114 1115 static int 1116 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 1117 const bus_dma_segment_t *dma_seg, int n_dma_seg, 1118 int vlan_tagged) 1119 { 1120 struct sfxge_tso_state tso; 1121 unsigned int id; 1122 unsigned skipped = 0; 1123 1124 tso_start(txq, &tso, dma_seg, mbuf); 1125 1126 while (dma_seg->ds_len + skipped <= tso.header_len) { 1127 skipped += dma_seg->ds_len; 1128 --n_dma_seg; 1129 KASSERT(n_dma_seg, ("no payload found in TSO packet")); 1130 ++dma_seg; 1131 } 1132 tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); 1133 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); 1134 1135 id = (txq->added + vlan_tagged) & txq->ptr_mask; 1136 if (__predict_false(tso_start_new_packet(txq, &tso, &id))) 1137 return (-1); 1138 1139 while (1) { 1140 tso_fill_packet_with_fragment(txq, &tso); 1141 /* Exactly one DMA descriptor is added */ 1142 KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); 1143 id = (id + 1) & txq->ptr_mask; 1144 1145 /* Move onto the next fragment? */ 1146 if (tso.in_len == 0) { 1147 --n_dma_seg; 1148 if (n_dma_seg == 0) 1149 break; 1150 ++dma_seg; 1151 tso.in_len = dma_seg->ds_len; 1152 tso.dma_addr = dma_seg->ds_addr; 1153 } 1154 1155 /* End of packet? */ 1156 if (tso.packet_space == 0) { 1157 /* If the queue is now full due to tiny MSS, 1158 * or we can't create another header, discard 1159 * the remainder of the input mbuf but do not 1160 * roll back the work we have done. 1161 */ 1162 if (txq->n_pend_desc + tso.fw_assisted + 1163 1 /* header */ + n_dma_seg > 1164 txq->max_pkt_desc) { 1165 txq->tso_pdrop_too_many++; 1166 break; 1167 } 1168 if (__predict_false(tso_start_new_packet(txq, &tso, 1169 &id))) { 1170 txq->tso_pdrop_no_rsrc++; 1171 break; 1172 } 1173 } 1174 } 1175 1176 txq->tso_bursts++; 1177 return (id); 1178 } 1179 1180 static void 1181 sfxge_tx_qunblock(struct sfxge_txq *txq) 1182 { 1183 struct sfxge_softc *sc; 1184 struct sfxge_evq *evq; 1185 1186 sc = txq->sc; 1187 evq = sc->evq[txq->evq_index]; 1188 1189 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 1190 1191 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) 1192 return; 1193 1194 SFXGE_TXQ_LOCK(txq); 1195 1196 if (txq->blocked) { 1197 unsigned int level; 1198 1199 level = txq->added - txq->completed; 1200 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) { 1201 /* reaped must be in sync with blocked */ 1202 sfxge_tx_qreap(txq); 1203 txq->blocked = 0; 1204 } 1205 } 1206 1207 sfxge_tx_qdpl_service(txq); 1208 /* note: lock has been dropped */ 1209 } 1210 1211 void 1212 sfxge_tx_qflush_done(struct sfxge_txq *txq) 1213 { 1214 1215 txq->flush_state = SFXGE_FLUSH_DONE; 1216 } 1217 1218 static void 1219 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) 1220 { 1221 struct sfxge_txq *txq; 1222 struct sfxge_evq *evq; 1223 unsigned int count; 1224 1225 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1226 1227 txq = sc->txq[index]; 1228 evq = sc->evq[txq->evq_index]; 1229 1230 SFXGE_EVQ_LOCK(evq); 1231 SFXGE_TXQ_LOCK(txq); 1232 1233 KASSERT(txq->init_state == SFXGE_TXQ_STARTED, 1234 ("txq->init_state != SFXGE_TXQ_STARTED")); 1235 1236 txq->init_state = SFXGE_TXQ_INITIALIZED; 1237 1238 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1239 txq->flush_state = SFXGE_FLUSH_PENDING; 1240 1241 SFXGE_EVQ_UNLOCK(evq); 1242 SFXGE_TXQ_UNLOCK(txq); 1243 1244 /* Flush the transmit queue. */ 1245 if (efx_tx_qflush(txq->common) != 0) { 1246 log(LOG_ERR, "%s: Flushing Tx queue %u failed\n", 1247 device_get_nameunit(sc->dev), index); 1248 txq->flush_state = SFXGE_FLUSH_DONE; 1249 } else { 1250 count = 0; 1251 do { 1252 /* Spin for 100ms. */ 1253 DELAY(100000); 1254 if (txq->flush_state != SFXGE_FLUSH_PENDING) 1255 break; 1256 } while (++count < 20); 1257 } 1258 SFXGE_EVQ_LOCK(evq); 1259 SFXGE_TXQ_LOCK(txq); 1260 1261 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, 1262 ("txq->flush_state == SFXGE_FLUSH_FAILED")); 1263 1264 if (txq->flush_state != SFXGE_FLUSH_DONE) { 1265 /* Flush timeout */ 1266 log(LOG_ERR, "%s: Cannot flush Tx queue %u\n", 1267 device_get_nameunit(sc->dev), index); 1268 txq->flush_state = SFXGE_FLUSH_DONE; 1269 } 1270 } 1271 1272 txq->blocked = 0; 1273 txq->pending = txq->added; 1274 1275 sfxge_tx_qcomplete(txq, evq); 1276 KASSERT(txq->completed == txq->added, 1277 ("txq->completed != txq->added")); 1278 1279 sfxge_tx_qreap(txq); 1280 KASSERT(txq->reaped == txq->completed, 1281 ("txq->reaped != txq->completed")); 1282 1283 txq->added = 0; 1284 txq->pending = 0; 1285 txq->completed = 0; 1286 txq->reaped = 0; 1287 1288 /* Destroy the common code transmit queue. */ 1289 efx_tx_qdestroy(txq->common); 1290 txq->common = NULL; 1291 1292 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1293 EFX_TXQ_NBUFS(sc->txq_entries)); 1294 1295 SFXGE_EVQ_UNLOCK(evq); 1296 SFXGE_TXQ_UNLOCK(txq); 1297 } 1298 1299 static int 1300 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) 1301 { 1302 struct sfxge_txq *txq; 1303 efsys_mem_t *esmp; 1304 uint16_t flags; 1305 struct sfxge_evq *evq; 1306 unsigned int desc_index; 1307 int rc; 1308 1309 SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); 1310 1311 txq = sc->txq[index]; 1312 esmp = &txq->mem; 1313 evq = sc->evq[txq->evq_index]; 1314 1315 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1316 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1317 KASSERT(evq->init_state == SFXGE_EVQ_STARTED, 1318 ("evq->init_state != SFXGE_EVQ_STARTED")); 1319 1320 /* Program the buffer table. */ 1321 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, 1322 EFX_TXQ_NBUFS(sc->txq_entries))) != 0) 1323 return (rc); 1324 1325 /* Determine the kind of queue we are creating. */ 1326 switch (txq->type) { 1327 case SFXGE_TXQ_NON_CKSUM: 1328 flags = 0; 1329 break; 1330 case SFXGE_TXQ_IP_CKSUM: 1331 flags = EFX_CKSUM_IPV4; 1332 break; 1333 case SFXGE_TXQ_IP_TCP_UDP_CKSUM: 1334 flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP; 1335 break; 1336 default: 1337 KASSERT(0, ("Impossible TX queue")); 1338 flags = 0; 1339 break; 1340 } 1341 1342 /* Create the common code transmit queue. */ 1343 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1344 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1345 &txq->common, &desc_index)) != 0) 1346 goto fail; 1347 1348 /* Initialise queue descriptor indexes */ 1349 txq->added = txq->pending = txq->completed = txq->reaped = desc_index; 1350 1351 SFXGE_TXQ_LOCK(txq); 1352 1353 /* Enable the transmit queue. */ 1354 efx_tx_qenable(txq->common); 1355 1356 txq->init_state = SFXGE_TXQ_STARTED; 1357 txq->flush_state = SFXGE_FLUSH_REQUIRED; 1358 1359 SFXGE_TXQ_UNLOCK(txq); 1360 1361 return (0); 1362 1363 fail: 1364 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1365 EFX_TXQ_NBUFS(sc->txq_entries)); 1366 return (rc); 1367 } 1368 1369 void 1370 sfxge_tx_stop(struct sfxge_softc *sc) 1371 { 1372 int index; 1373 1374 index = sc->txq_count; 1375 while (--index >= 0) 1376 sfxge_tx_qstop(sc, index); 1377 1378 /* Tear down the transmit module */ 1379 efx_tx_fini(sc->enp); 1380 } 1381 1382 int 1383 sfxge_tx_start(struct sfxge_softc *sc) 1384 { 1385 int index; 1386 int rc; 1387 1388 /* Initialize the common code transmit module. */ 1389 if ((rc = efx_tx_init(sc->enp)) != 0) 1390 return (rc); 1391 1392 for (index = 0; index < sc->txq_count; index++) { 1393 if ((rc = sfxge_tx_qstart(sc, index)) != 0) 1394 goto fail; 1395 } 1396 1397 return (0); 1398 1399 fail: 1400 while (--index >= 0) 1401 sfxge_tx_qstop(sc, index); 1402 1403 efx_tx_fini(sc->enp); 1404 1405 return (rc); 1406 } 1407 1408 static int 1409 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node) 1410 { 1411 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev); 1412 struct sysctl_oid *stat_node; 1413 unsigned int id; 1414 1415 stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1416 "stats", CTLFLAG_RD, NULL, 1417 "Tx queue statistics"); 1418 if (stat_node == NULL) 1419 return (ENOMEM); 1420 1421 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1422 SYSCTL_ADD_ULONG( 1423 ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO, 1424 sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS, 1425 (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset), 1426 ""); 1427 } 1428 1429 return (0); 1430 } 1431 1432 /** 1433 * Destroy a transmit queue. 1434 */ 1435 static void 1436 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) 1437 { 1438 struct sfxge_txq *txq; 1439 unsigned int nmaps; 1440 1441 txq = sc->txq[index]; 1442 1443 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1444 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1445 1446 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) 1447 tso_fini(txq); 1448 1449 /* Free the context arrays. */ 1450 free(txq->pend_desc, M_SFXGE); 1451 nmaps = sc->txq_entries; 1452 while (nmaps-- != 0) 1453 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1454 free(txq->stmp, M_SFXGE); 1455 1456 /* Release DMA memory mapping. */ 1457 sfxge_dma_free(&txq->mem); 1458 1459 sc->txq[index] = NULL; 1460 1461 SFXGE_TXQ_LOCK_DESTROY(txq); 1462 1463 free(txq, M_SFXGE); 1464 } 1465 1466 /* 1467 * Estimate maximum number of Tx descriptors required for TSO packet. 1468 * With minimum MSS and maximum mbuf length we might need more (even 1469 * than a ring-ful of descriptors), but this should not happen in 1470 * practice except due to deliberate attack. In that case we will 1471 * truncate the output at a packet boundary. 1472 */ 1473 static unsigned int 1474 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type) 1475 { 1476 /* One descriptor for every input fragment */ 1477 unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG; 1478 1479 /* VLAN tagging Tx option descriptor may be required */ 1480 if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled) 1481 max_descs++; 1482 1483 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) { 1484 /* 1485 * Plus header and payload descriptor for each output segment. 1486 * Minus one since header fragment is already counted. 1487 */ 1488 max_descs += SFXGE_TSO_MAX_SEGS * 2 - 1; 1489 1490 /* FW assisted TSO requires one more descriptor per segment */ 1491 if (sc->tso_fw_assisted) 1492 max_descs += SFXGE_TSO_MAX_SEGS; 1493 } 1494 1495 return (max_descs); 1496 } 1497 1498 static int 1499 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, 1500 enum sfxge_txq_type type, unsigned int evq_index) 1501 { 1502 char name[16]; 1503 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1504 struct sysctl_oid *txq_node; 1505 struct sfxge_txq *txq; 1506 struct sfxge_evq *evq; 1507 struct sfxge_tx_dpl *stdp; 1508 struct sysctl_oid *dpl_node; 1509 efsys_mem_t *esmp; 1510 unsigned int nmaps; 1511 int rc; 1512 1513 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); 1514 txq->sc = sc; 1515 txq->entries = sc->txq_entries; 1516 txq->ptr_mask = txq->entries - 1; 1517 1518 sc->txq[txq_index] = txq; 1519 esmp = &txq->mem; 1520 1521 evq = sc->evq[evq_index]; 1522 1523 /* Allocate and zero DMA space for the descriptor ring. */ 1524 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) 1525 return (rc); 1526 1527 /* Allocate buffer table entries. */ 1528 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), 1529 &txq->buf_base_id); 1530 1531 /* Create a DMA tag for packet mappings. */ 1532 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000, 1533 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, 1534 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL, 1535 &txq->packet_dma_tag) != 0) { 1536 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); 1537 rc = ENOMEM; 1538 goto fail; 1539 } 1540 1541 /* Allocate pending descriptor array for batching writes. */ 1542 txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries, 1543 M_SFXGE, M_ZERO | M_WAITOK); 1544 1545 /* Allocate and initialise mbuf DMA mapping array. */ 1546 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, 1547 M_SFXGE, M_ZERO | M_WAITOK); 1548 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { 1549 rc = bus_dmamap_create(txq->packet_dma_tag, 0, 1550 &txq->stmp[nmaps].map); 1551 if (rc != 0) 1552 goto fail2; 1553 } 1554 1555 snprintf(name, sizeof(name), "%u", txq_index); 1556 txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node), 1557 OID_AUTO, name, CTLFLAG_RD, NULL, ""); 1558 if (txq_node == NULL) { 1559 rc = ENOMEM; 1560 goto fail_txq_node; 1561 } 1562 1563 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && 1564 (rc = tso_init(txq)) != 0) 1565 goto fail3; 1566 1567 if (sfxge_tx_dpl_get_max <= 0) { 1568 log(LOG_ERR, "%s=%d must be greater than 0", 1569 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); 1570 rc = EINVAL; 1571 goto fail_tx_dpl_get_max; 1572 } 1573 if (sfxge_tx_dpl_get_non_tcp_max <= 0) { 1574 log(LOG_ERR, "%s=%d must be greater than 0", 1575 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, 1576 sfxge_tx_dpl_get_non_tcp_max); 1577 rc = EINVAL; 1578 goto fail_tx_dpl_get_max; 1579 } 1580 if (sfxge_tx_dpl_put_max < 0) { 1581 log(LOG_ERR, "%s=%d must be greater or equal to 0", 1582 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); 1583 rc = EINVAL; 1584 goto fail_tx_dpl_put_max; 1585 } 1586 1587 /* Initialize the deferred packet list. */ 1588 stdp = &txq->dpl; 1589 stdp->std_put_max = sfxge_tx_dpl_put_max; 1590 stdp->std_get_max = sfxge_tx_dpl_get_max; 1591 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; 1592 stdp->std_getp = &stdp->std_get; 1593 1594 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index); 1595 1596 dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1597 "dpl", CTLFLAG_RD, NULL, 1598 "Deferred packet list statistics"); 1599 if (dpl_node == NULL) { 1600 rc = ENOMEM; 1601 goto fail_dpl_node; 1602 } 1603 1604 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1605 "get_count", CTLFLAG_RD | CTLFLAG_STATS, 1606 &stdp->std_get_count, 0, ""); 1607 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1608 "get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, 1609 &stdp->std_get_non_tcp_count, 0, ""); 1610 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1611 "get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1612 &stdp->std_get_hiwat, 0, ""); 1613 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1614 "put_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1615 &stdp->std_put_hiwat, 0, ""); 1616 1617 rc = sfxge_txq_stat_init(txq, txq_node); 1618 if (rc != 0) 1619 goto fail_txq_stat_init; 1620 1621 txq->type = type; 1622 txq->evq_index = evq_index; 1623 txq->txq_index = txq_index; 1624 txq->init_state = SFXGE_TXQ_INITIALIZED; 1625 txq->hw_vlan_tci = 0; 1626 1627 txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, type); 1628 1629 return (0); 1630 1631 fail_txq_stat_init: 1632 fail_dpl_node: 1633 fail_tx_dpl_put_max: 1634 fail_tx_dpl_get_max: 1635 fail3: 1636 fail_txq_node: 1637 free(txq->pend_desc, M_SFXGE); 1638 fail2: 1639 while (nmaps-- != 0) 1640 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1641 free(txq->stmp, M_SFXGE); 1642 bus_dma_tag_destroy(txq->packet_dma_tag); 1643 1644 fail: 1645 sfxge_dma_free(esmp); 1646 1647 return (rc); 1648 } 1649 1650 static int 1651 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS) 1652 { 1653 struct sfxge_softc *sc = arg1; 1654 unsigned int id = arg2; 1655 unsigned long sum; 1656 unsigned int index; 1657 1658 /* Sum across all TX queues */ 1659 sum = 0; 1660 for (index = 0; index < sc->txq_count; index++) 1661 sum += *(unsigned long *)((caddr_t)sc->txq[index] + 1662 sfxge_tx_stats[id].offset); 1663 1664 return (SYSCTL_OUT(req, &sum, sizeof(sum))); 1665 } 1666 1667 static void 1668 sfxge_tx_stat_init(struct sfxge_softc *sc) 1669 { 1670 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1671 struct sysctl_oid_list *stat_list; 1672 unsigned int id; 1673 1674 stat_list = SYSCTL_CHILDREN(sc->stats_node); 1675 1676 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1677 SYSCTL_ADD_PROC( 1678 ctx, stat_list, 1679 OID_AUTO, sfxge_tx_stats[id].name, 1680 CTLTYPE_ULONG|CTLFLAG_RD, 1681 sc, id, sfxge_tx_stat_handler, "LU", 1682 ""); 1683 } 1684 } 1685 1686 uint64_t 1687 sfxge_tx_get_drops(struct sfxge_softc *sc) 1688 { 1689 unsigned int index; 1690 uint64_t drops = 0; 1691 struct sfxge_txq *txq; 1692 1693 /* Sum across all TX queues */ 1694 for (index = 0; index < sc->txq_count; index++) { 1695 txq = sc->txq[index]; 1696 /* 1697 * In theory, txq->put_overflow and txq->netdown_drops 1698 * should use atomic operation and other should be 1699 * obtained under txq lock, but it is just statistics. 1700 */ 1701 drops += txq->drops + txq->get_overflow + 1702 txq->get_non_tcp_overflow + 1703 txq->put_overflow + txq->netdown_drops + 1704 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc; 1705 } 1706 return (drops); 1707 } 1708 1709 void 1710 sfxge_tx_fini(struct sfxge_softc *sc) 1711 { 1712 int index; 1713 1714 index = sc->txq_count; 1715 while (--index >= 0) 1716 sfxge_tx_qfini(sc, index); 1717 1718 sc->txq_count = 0; 1719 } 1720 1721 1722 int 1723 sfxge_tx_init(struct sfxge_softc *sc) 1724 { 1725 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); 1726 struct sfxge_intr *intr; 1727 int index; 1728 int rc; 1729 1730 intr = &sc->intr; 1731 1732 KASSERT(intr->state == SFXGE_INTR_INITIALIZED, 1733 ("intr->state != SFXGE_INTR_INITIALIZED")); 1734 1735 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc; 1736 1737 sc->tso_fw_assisted = sfxge_tso_fw_assisted; 1738 if (sc->tso_fw_assisted) 1739 sc->tso_fw_assisted = 1740 (encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) && 1741 (encp->enc_fw_assisted_tso_enabled); 1742 1743 sc->txqs_node = SYSCTL_ADD_NODE( 1744 device_get_sysctl_ctx(sc->dev), 1745 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), 1746 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); 1747 if (sc->txqs_node == NULL) { 1748 rc = ENOMEM; 1749 goto fail_txq_node; 1750 } 1751 1752 /* Initialize the transmit queues */ 1753 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, 1754 SFXGE_TXQ_NON_CKSUM, 0)) != 0) 1755 goto fail; 1756 1757 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, 1758 SFXGE_TXQ_IP_CKSUM, 0)) != 0) 1759 goto fail2; 1760 1761 for (index = 0; 1762 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1; 1763 index++) { 1764 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index, 1765 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) 1766 goto fail3; 1767 } 1768 1769 sfxge_tx_stat_init(sc); 1770 1771 return (0); 1772 1773 fail3: 1774 while (--index >= 0) 1775 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1776 1777 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 1778 1779 fail2: 1780 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 1781 1782 fail: 1783 fail_txq_node: 1784 sc->txq_count = 0; 1785 return (rc); 1786 } 1787