sfxge_tx.c revision 281955
1205147Sedwin/*- 2205147Sedwin * Copyright (c) 2010-2011 Solarflare Communications, Inc. 3205147Sedwin * All rights reserved. 4205147Sedwin * 5205147Sedwin * This software was developed in part by Philip Paeps under contract for 6205147Sedwin * Solarflare Communications, Inc. 7205147Sedwin * 8205147Sedwin * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30/* Theory of operation: 31 * 32 * Tx queues allocation and mapping 33 * 34 * One Tx queue with enabled checksum offload is allocated per Rx channel 35 * (event queue). Also 2 Tx queues (one without checksum offload and one 36 * with IP checksum offload only) are allocated and bound to event queue 0. 37 * sfxge_txq_type is used as Tx queue label. 38 * 39 * So, event queue plus label mapping to Tx queue index is: 40 * if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES) 41 * else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1 42 * See sfxge_get_txq_by_label() sfxge_ev.c 43 */ 44 45#include <sys/cdefs.h> 46__FBSDID("$FreeBSD: stable/10/sys/dev/sfxge/sfxge_tx.c 281955 2015-04-24 23:26:44Z hiren $"); 47 48#include <sys/types.h> 49#include <sys/mbuf.h> 50#include <sys/smp.h> 51#include <sys/socket.h> 52#include <sys/sysctl.h> 53#include <sys/syslog.h> 54 55#include <net/bpf.h> 56#include <net/ethernet.h> 57#include <net/if.h> 58#include <net/if_vlan_var.h> 59 60#include <netinet/in.h> 61#include <netinet/ip.h> 62#include <netinet/ip6.h> 63#include <netinet/tcp.h> 64 65#include "common/efx.h" 66 67#include "sfxge.h" 68#include "sfxge_tx.h" 69 70/* 71 * Estimate maximum number of Tx descriptors required for TSO packet. 72 * With minimum MSS and maximum mbuf length we might need more (even 73 * than a ring-ful of descriptors), but this should not happen in 74 * practice except due to deliberate attack. In that case we will 75 * truncate the output at a packet boundary. 76 */ 77#define SFXGE_TSO_MAX_DESC \ 78 (SFXGE_TSO_MAX_SEGS * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1) 79 80/* 81 * Set the block level to ensure there is space to generate a 82 * large number of descriptors for TSO. 83 */ 84#define SFXGE_TXQ_BLOCK_LEVEL(_entries) \ 85 (EFX_TXQ_LIMIT(_entries) - SFXGE_TSO_MAX_DESC) 86 87 88#define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max) 89static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT; 90TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max); 91SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN, 92 &sfxge_tx_dpl_get_max, 0, 93 "Maximum number of any packets in deferred packet get-list"); 94 95#define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \ 96 SFXGE_PARAM(tx_dpl_get_non_tcp_max) 97static int sfxge_tx_dpl_get_non_tcp_max = 98 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT; 99TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max); 100SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN, 101 &sfxge_tx_dpl_get_non_tcp_max, 0, 102 "Maximum number of non-TCP packets in deferred packet get-list"); 103 104#define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max) 105static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT; 106TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max); 107SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN, 108 &sfxge_tx_dpl_put_max, 0, 109 "Maximum number of any packets in deferred packet put-list"); 110 111 112static const struct { 113 const char *name; 114 size_t offset; 115} sfxge_tx_stats[] = { 116#define SFXGE_TX_STAT(name, member) \ 117 { #name, offsetof(struct sfxge_txq, member) } 118 SFXGE_TX_STAT(tso_bursts, tso_bursts), 119 SFXGE_TX_STAT(tso_packets, tso_packets), 120 SFXGE_TX_STAT(tso_long_headers, tso_long_headers), 121 SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many), 122 SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc), 123 SFXGE_TX_STAT(tx_collapses, collapses), 124 SFXGE_TX_STAT(tx_drops, drops), 125 SFXGE_TX_STAT(tx_get_overflow, get_overflow), 126 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow), 127 SFXGE_TX_STAT(tx_put_overflow, put_overflow), 128 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops), 129}; 130 131 132/* Forward declarations. */ 133static void sfxge_tx_qdpl_service(struct sfxge_txq *txq); 134static void sfxge_tx_qlist_post(struct sfxge_txq *txq); 135static void sfxge_tx_qunblock(struct sfxge_txq *txq); 136static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 137 const bus_dma_segment_t *dma_seg, int n_dma_seg); 138 139void 140sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) 141{ 142 unsigned int completed; 143 144 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 145 146 completed = txq->completed; 147 while (completed != txq->pending) { 148 struct sfxge_tx_mapping *stmp; 149 unsigned int id; 150 151 id = completed++ & txq->ptr_mask; 152 153 stmp = &txq->stmp[id]; 154 if (stmp->flags & TX_BUF_UNMAP) { 155 bus_dmamap_unload(txq->packet_dma_tag, stmp->map); 156 if (stmp->flags & TX_BUF_MBUF) { 157 struct mbuf *m = stmp->u.mbuf; 158 do 159 m = m_free(m); 160 while (m != NULL); 161 } else { 162 free(stmp->u.heap_buf, M_SFXGE); 163 } 164 stmp->flags = 0; 165 } 166 } 167 txq->completed = completed; 168 169 /* Check whether we need to unblock the queue. */ 170 mb(); 171 if (txq->blocked) { 172 unsigned int level; 173 174 level = txq->added - txq->completed; 175 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) 176 sfxge_tx_qunblock(txq); 177 } 178} 179 180static unsigned int 181sfxge_is_mbuf_non_tcp(struct mbuf *mbuf) 182{ 183 /* Absense of TCP checksum flags does not mean that it is non-TCP 184 * but it should be true if user wants to achieve high throughput. 185 */ 186 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP))); 187} 188 189/* 190 * Reorder the put list and append it to the get list. 191 */ 192static void 193sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq) 194{ 195 struct sfxge_tx_dpl *stdp; 196 struct mbuf *mbuf, *get_next, **get_tailp; 197 volatile uintptr_t *putp; 198 uintptr_t put; 199 unsigned int count; 200 unsigned int non_tcp_count; 201 202 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 203 204 stdp = &txq->dpl; 205 206 /* Acquire the put list. */ 207 putp = &stdp->std_put; 208 put = atomic_readandclear_ptr(putp); 209 mbuf = (void *)put; 210 211 if (mbuf == NULL) 212 return; 213 214 /* Reverse the put list. */ 215 get_tailp = &mbuf->m_nextpkt; 216 get_next = NULL; 217 218 count = 0; 219 non_tcp_count = 0; 220 do { 221 struct mbuf *put_next; 222 223 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf); 224 put_next = mbuf->m_nextpkt; 225 mbuf->m_nextpkt = get_next; 226 get_next = mbuf; 227 mbuf = put_next; 228 229 count++; 230 } while (mbuf != NULL); 231 232 if (count > stdp->std_put_hiwat) 233 stdp->std_put_hiwat = count; 234 235 /* Append the reversed put list to the get list. */ 236 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL")); 237 *stdp->std_getp = get_next; 238 stdp->std_getp = get_tailp; 239 stdp->std_get_count += count; 240 stdp->std_get_non_tcp_count += non_tcp_count; 241} 242 243static void 244sfxge_tx_qreap(struct sfxge_txq *txq) 245{ 246 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 247 248 txq->reaped = txq->completed; 249} 250 251static void 252sfxge_tx_qlist_post(struct sfxge_txq *txq) 253{ 254 unsigned int old_added; 255 unsigned int level; 256 int rc; 257 258 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 259 260 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0")); 261 KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC, 262 ("txq->n_pend_desc too large")); 263 KASSERT(!txq->blocked, ("txq->blocked")); 264 265 old_added = txq->added; 266 267 /* Post the fragment list. */ 268 rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc, 269 txq->reaped, &txq->added); 270 KASSERT(rc == 0, ("efx_tx_qpost() failed")); 271 272 /* If efx_tx_qpost() had to refragment, our information about 273 * buffers to free may be associated with the wrong 274 * descriptors. 275 */ 276 KASSERT(txq->added - old_added == txq->n_pend_desc, 277 ("efx_tx_qpost() refragmented descriptors")); 278 279 level = txq->added - txq->reaped; 280 KASSERT(level <= txq->entries, ("overfilled TX queue")); 281 282 /* Clear the fragment list. */ 283 txq->n_pend_desc = 0; 284 285 /* Have we reached the block level? */ 286 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) 287 return; 288 289 /* Reap, and check again */ 290 sfxge_tx_qreap(txq); 291 level = txq->added - txq->reaped; 292 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) 293 return; 294 295 txq->blocked = 1; 296 297 /* 298 * Avoid a race with completion interrupt handling that could leave 299 * the queue blocked. 300 */ 301 mb(); 302 sfxge_tx_qreap(txq); 303 level = txq->added - txq->reaped; 304 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) { 305 mb(); 306 txq->blocked = 0; 307 } 308} 309 310static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf) 311{ 312 bus_dmamap_t *used_map; 313 bus_dmamap_t map; 314 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG]; 315 unsigned int id; 316 struct sfxge_tx_mapping *stmp; 317 efx_buffer_t *desc; 318 int n_dma_seg; 319 int rc; 320 int i; 321 322 KASSERT(!txq->blocked, ("txq->blocked")); 323 324 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) 325 prefetch_read_many(mbuf->m_data); 326 327 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) { 328 rc = EINTR; 329 goto reject; 330 } 331 332 /* Load the packet for DMA. */ 333 id = txq->added & txq->ptr_mask; 334 stmp = &txq->stmp[id]; 335 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, 336 mbuf, dma_seg, &n_dma_seg, 0); 337 if (rc == EFBIG) { 338 /* Try again. */ 339 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, 340 SFXGE_TX_MAPPING_MAX_SEG); 341 if (new_mbuf == NULL) 342 goto reject; 343 ++txq->collapses; 344 mbuf = new_mbuf; 345 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, 346 stmp->map, mbuf, 347 dma_seg, &n_dma_seg, 0); 348 } 349 if (rc != 0) 350 goto reject; 351 352 /* Make the packet visible to the hardware. */ 353 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); 354 355 used_map = &stmp->map; 356 357 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 358 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg); 359 if (rc < 0) 360 goto reject_mapped; 361 stmp = &txq->stmp[rc]; 362 } else { 363 /* Add the mapping to the fragment list, and set flags 364 * for the buffer. 365 */ 366 i = 0; 367 for (;;) { 368 desc = &txq->pend_desc[i]; 369 desc->eb_addr = dma_seg[i].ds_addr; 370 desc->eb_size = dma_seg[i].ds_len; 371 if (i == n_dma_seg - 1) { 372 desc->eb_eop = 1; 373 break; 374 } 375 desc->eb_eop = 0; 376 i++; 377 378 stmp->flags = 0; 379 if (__predict_false(stmp == 380 &txq->stmp[txq->ptr_mask])) 381 stmp = &txq->stmp[0]; 382 else 383 stmp++; 384 } 385 txq->n_pend_desc = n_dma_seg; 386 } 387 388 /* 389 * If the mapping required more than one descriptor 390 * then we need to associate the DMA map with the last 391 * descriptor, not the first. 392 */ 393 if (used_map != &stmp->map) { 394 map = stmp->map; 395 stmp->map = *used_map; 396 *used_map = map; 397 } 398 399 stmp->u.mbuf = mbuf; 400 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; 401 402 /* Post the fragment list. */ 403 sfxge_tx_qlist_post(txq); 404 405 return (0); 406 407reject_mapped: 408 bus_dmamap_unload(txq->packet_dma_tag, *used_map); 409reject: 410 /* Drop the packet on the floor. */ 411 m_freem(mbuf); 412 ++txq->drops; 413 414 return (rc); 415} 416 417/* 418 * Drain the deferred packet list into the transmit queue. 419 */ 420static void 421sfxge_tx_qdpl_drain(struct sfxge_txq *txq) 422{ 423 struct sfxge_softc *sc; 424 struct sfxge_tx_dpl *stdp; 425 struct mbuf *mbuf, *next; 426 unsigned int count; 427 unsigned int non_tcp_count; 428 unsigned int pushed; 429 int rc; 430 431 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 432 433 sc = txq->sc; 434 stdp = &txq->dpl; 435 pushed = txq->added; 436 437 if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) { 438 prefetch_read_many(sc->enp); 439 prefetch_read_many(txq->common); 440 } 441 442 mbuf = stdp->std_get; 443 count = stdp->std_get_count; 444 non_tcp_count = stdp->std_get_non_tcp_count; 445 446 if (count > stdp->std_get_hiwat) 447 stdp->std_get_hiwat = count; 448 449 while (count != 0) { 450 KASSERT(mbuf != NULL, ("mbuf == NULL")); 451 452 next = mbuf->m_nextpkt; 453 mbuf->m_nextpkt = NULL; 454 455 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ 456 457 if (next != NULL) 458 prefetch_read_many(next); 459 460 rc = sfxge_tx_queue_mbuf(txq, mbuf); 461 --count; 462 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); 463 mbuf = next; 464 if (rc != 0) 465 continue; 466 467 if (txq->blocked) 468 break; 469 470 /* Push the fragments to the hardware in batches. */ 471 if (txq->added - pushed >= SFXGE_TX_BATCH) { 472 efx_tx_qpush(txq->common, txq->added); 473 pushed = txq->added; 474 } 475 } 476 477 if (count == 0) { 478 KASSERT(mbuf == NULL, ("mbuf != NULL")); 479 KASSERT(non_tcp_count == 0, 480 ("inconsistent TCP/non-TCP detection")); 481 stdp->std_get = NULL; 482 stdp->std_get_count = 0; 483 stdp->std_get_non_tcp_count = 0; 484 stdp->std_getp = &stdp->std_get; 485 } else { 486 stdp->std_get = mbuf; 487 stdp->std_get_count = count; 488 stdp->std_get_non_tcp_count = non_tcp_count; 489 } 490 491 if (txq->added != pushed) 492 efx_tx_qpush(txq->common, txq->added); 493 494 KASSERT(txq->blocked || stdp->std_get_count == 0, 495 ("queue unblocked but count is non-zero")); 496} 497 498#define SFXGE_TX_QDPL_PENDING(_txq) \ 499 ((_txq)->dpl.std_put != 0) 500 501/* 502 * Service the deferred packet list. 503 * 504 * NOTE: drops the txq mutex! 505 */ 506static void 507sfxge_tx_qdpl_service(struct sfxge_txq *txq) 508{ 509 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 510 511 do { 512 if (SFXGE_TX_QDPL_PENDING(txq)) 513 sfxge_tx_qdpl_swizzle(txq); 514 515 if (!txq->blocked) 516 sfxge_tx_qdpl_drain(txq); 517 518 SFXGE_TXQ_UNLOCK(txq); 519 } while (SFXGE_TX_QDPL_PENDING(txq) && 520 SFXGE_TXQ_TRYLOCK(txq)); 521} 522 523/* 524 * Put a packet on the deferred packet list. 525 * 526 * If we are called with the txq lock held, we put the packet on the "get 527 * list", otherwise we atomically push it on the "put list". The swizzle 528 * function takes care of ordering. 529 * 530 * The length of the put list is bounded by SFXGE_TX_MAX_DEFERRED. We 531 * overload the csum_data field in the mbuf to keep track of this length 532 * because there is no cheap alternative to avoid races. 533 */ 534static int 535sfxge_tx_qdpl_put(struct sfxge_txq *txq, struct mbuf *mbuf, int locked) 536{ 537 struct sfxge_tx_dpl *stdp; 538 539 stdp = &txq->dpl; 540 541 KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); 542 543 if (locked) { 544 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); 545 546 sfxge_tx_qdpl_swizzle(txq); 547 548 if (stdp->std_get_count >= stdp->std_get_max) { 549 txq->get_overflow++; 550 return (ENOBUFS); 551 } 552 if (sfxge_is_mbuf_non_tcp(mbuf)) { 553 if (stdp->std_get_non_tcp_count >= 554 stdp->std_get_non_tcp_max) { 555 txq->get_non_tcp_overflow++; 556 return (ENOBUFS); 557 } 558 stdp->std_get_non_tcp_count++; 559 } 560 561 *(stdp->std_getp) = mbuf; 562 stdp->std_getp = &mbuf->m_nextpkt; 563 stdp->std_get_count++; 564 } else { 565 volatile uintptr_t *putp; 566 uintptr_t old; 567 uintptr_t new; 568 unsigned old_len; 569 570 putp = &stdp->std_put; 571 new = (uintptr_t)mbuf; 572 573 do { 574 old = *putp; 575 if (old != 0) { 576 struct mbuf *mp = (struct mbuf *)old; 577 old_len = mp->m_pkthdr.csum_data; 578 } else 579 old_len = 0; 580 if (old_len >= stdp->std_put_max) { 581 atomic_add_long(&txq->put_overflow, 1); 582 return (ENOBUFS); 583 } 584 mbuf->m_pkthdr.csum_data = old_len + 1; 585 mbuf->m_nextpkt = (void *)old; 586 } while (atomic_cmpset_ptr(putp, old, new) == 0); 587 } 588 589 return (0); 590} 591 592/* 593 * Called from if_transmit - will try to grab the txq lock and enqueue to the 594 * put list if it succeeds, otherwise try to push onto the defer list if space. 595 */ 596int 597sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) 598{ 599 int locked; 600 int rc; 601 602 if (!SFXGE_LINK_UP(txq->sc)) { 603 rc = ENETDOWN; 604 atomic_add_long(&txq->netdown_drops, 1); 605 goto fail; 606 } 607 608 /* 609 * Try to grab the txq lock. If we are able to get the lock, 610 * the packet will be appended to the "get list" of the deferred 611 * packet list. Otherwise, it will be pushed on the "put list". 612 */ 613 locked = SFXGE_TXQ_TRYLOCK(txq); 614 615 if (sfxge_tx_qdpl_put(txq, m, locked) != 0) { 616 if (locked) 617 SFXGE_TXQ_UNLOCK(txq); 618 rc = ENOBUFS; 619 goto fail; 620 } 621 622 /* 623 * Try to grab the lock again. 624 * 625 * If we are able to get the lock, we need to process the deferred 626 * packet list. If we are not able to get the lock, another thread 627 * is processing the list. 628 */ 629 if (!locked) 630 locked = SFXGE_TXQ_TRYLOCK(txq); 631 632 if (locked) { 633 /* Try to service the list. */ 634 sfxge_tx_qdpl_service(txq); 635 /* Lock has been dropped. */ 636 } 637 638 return (0); 639 640fail: 641 m_freem(m); 642 return (rc); 643} 644 645static void 646sfxge_tx_qdpl_flush(struct sfxge_txq *txq) 647{ 648 struct sfxge_tx_dpl *stdp = &txq->dpl; 649 struct mbuf *mbuf, *next; 650 651 SFXGE_TXQ_LOCK(txq); 652 653 sfxge_tx_qdpl_swizzle(txq); 654 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { 655 next = mbuf->m_nextpkt; 656 m_freem(mbuf); 657 } 658 stdp->std_get = NULL; 659 stdp->std_get_count = 0; 660 stdp->std_get_non_tcp_count = 0; 661 stdp->std_getp = &stdp->std_get; 662 663 SFXGE_TXQ_UNLOCK(txq); 664} 665 666void 667sfxge_if_qflush(struct ifnet *ifp) 668{ 669 struct sfxge_softc *sc; 670 unsigned int i; 671 672 sc = ifp->if_softc; 673 674 for (i = 0; i < sc->txq_count; i++) 675 sfxge_tx_qdpl_flush(sc->txq[i]); 676} 677 678/* 679 * TX start -- called by the stack. 680 */ 681int 682sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) 683{ 684 struct sfxge_softc *sc; 685 struct sfxge_txq *txq; 686 int rc; 687 688 sc = (struct sfxge_softc *)ifp->if_softc; 689 690 /* 691 * Transmit may be called when interface is up from the kernel 692 * point of view, but not yet up (in progress) from the driver 693 * point of view. I.e. link aggregation bring up. 694 * Transmit may be called when interface is up from the driver 695 * point of view, but already down from the kernel point of 696 * view. I.e. Rx when interface shutdown is in progress. 697 */ 698 KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP), 699 ("interface not up")); 700 701 /* Pick the desired transmit queue. */ 702 if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) { 703 int index = 0; 704 705 /* check if flowid is set */ 706 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 707 uint32_t hash = m->m_pkthdr.flowid; 708 709 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX]; 710 } 711 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; 712 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { 713 txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; 714 } else { 715 txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; 716 } 717 718 rc = sfxge_tx_packet_add(txq, m); 719 720 return (rc); 721} 722 723/* 724 * Software "TSO". Not quite as good as doing it in hardware, but 725 * still faster than segmenting in the stack. 726 */ 727 728struct sfxge_tso_state { 729 /* Output position */ 730 unsigned out_len; /* Remaining length in current segment */ 731 unsigned seqnum; /* Current sequence number */ 732 unsigned packet_space; /* Remaining space in current packet */ 733 734 /* Input position */ 735 uint64_t dma_addr; /* DMA address of current position */ 736 unsigned in_len; /* Remaining length in current mbuf */ 737 738 const struct mbuf *mbuf; /* Input mbuf (head of chain) */ 739 u_short protocol; /* Network protocol (after VLAN decap) */ 740 ssize_t nh_off; /* Offset of network header */ 741 ssize_t tcph_off; /* Offset of TCP header */ 742 unsigned header_len; /* Number of bytes of header */ 743 unsigned seg_size; /* TCP segment size */ 744}; 745 746static const struct ip *tso_iph(const struct sfxge_tso_state *tso) 747{ 748 KASSERT(tso->protocol == htons(ETHERTYPE_IP), 749 ("tso_iph() in non-IPv4 state")); 750 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); 751} 752static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) 753{ 754 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 755 ("tso_ip6h() in non-IPv6 state")); 756 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); 757} 758static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) 759{ 760 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); 761} 762 763/* Size of preallocated TSO header buffers. Larger blocks must be 764 * allocated from the heap. 765 */ 766#define TSOH_STD_SIZE 128 767 768/* At most half the descriptors in the queue at any time will refer to 769 * a TSO header buffer, since they must always be followed by a 770 * payload descriptor referring to an mbuf. 771 */ 772#define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) 773#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 774#define TSOH_PAGE_COUNT(_txq_entries) \ 775 ((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE) 776 777static int tso_init(struct sfxge_txq *txq) 778{ 779 struct sfxge_softc *sc = txq->sc; 780 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); 781 int i, rc; 782 783 /* Allocate TSO header buffers */ 784 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), 785 M_SFXGE, M_WAITOK); 786 787 for (i = 0; i < tsoh_page_count; i++) { 788 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); 789 if (rc != 0) 790 goto fail; 791 } 792 793 return (0); 794 795fail: 796 while (i-- > 0) 797 sfxge_dma_free(&txq->tsoh_buffer[i]); 798 free(txq->tsoh_buffer, M_SFXGE); 799 txq->tsoh_buffer = NULL; 800 return (rc); 801} 802 803static void tso_fini(struct sfxge_txq *txq) 804{ 805 int i; 806 807 if (txq->tsoh_buffer != NULL) { 808 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) 809 sfxge_dma_free(&txq->tsoh_buffer[i]); 810 free(txq->tsoh_buffer, M_SFXGE); 811 } 812} 813 814static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf) 815{ 816 struct ether_header *eh = mtod(mbuf, struct ether_header *); 817 const struct tcphdr *th; 818 struct tcphdr th_copy; 819 820 tso->mbuf = mbuf; 821 822 /* Find network protocol and header */ 823 tso->protocol = eh->ether_type; 824 if (tso->protocol == htons(ETHERTYPE_VLAN)) { 825 struct ether_vlan_header *veh = 826 mtod(mbuf, struct ether_vlan_header *); 827 tso->protocol = veh->evl_proto; 828 tso->nh_off = sizeof(*veh); 829 } else { 830 tso->nh_off = sizeof(*eh); 831 } 832 833 /* Find TCP header */ 834 if (tso->protocol == htons(ETHERTYPE_IP)) { 835 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, 836 ("TSO required on non-TCP packet")); 837 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; 838 } else { 839 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), 840 ("TSO required on non-IP packet")); 841 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, 842 ("TSO required on non-TCP packet")); 843 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); 844 } 845 846 KASSERT(mbuf->m_len >= tso->tcph_off, 847 ("network header is fragmented in mbuf")); 848 /* We need TCP header including flags (window is the next) */ 849 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) { 850 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy), 851 (caddr_t)&th_copy); 852 th = &th_copy; 853 } else { 854 th = tso_tcph(tso); 855 } 856 857 tso->header_len = tso->tcph_off + 4 * th->th_off; 858 tso->seg_size = mbuf->m_pkthdr.tso_segsz; 859 860 tso->seqnum = ntohl(th->th_seq); 861 862 /* These flags must not be duplicated */ 863 KASSERT(!(th->th_flags & (TH_URG | TH_SYN | TH_RST)), 864 ("incompatible TCP flag on TSO packet")); 865 866 tso->out_len = mbuf->m_pkthdr.len - tso->header_len; 867} 868 869/* 870 * tso_fill_packet_with_fragment - form descriptors for the current fragment 871 * 872 * Form descriptors for the current fragment, until we reach the end 873 * of fragment or end-of-packet. Return 0 on success, 1 if not enough 874 * space. 875 */ 876static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, 877 struct sfxge_tso_state *tso) 878{ 879 efx_buffer_t *desc; 880 int n; 881 882 if (tso->in_len == 0 || tso->packet_space == 0) 883 return; 884 885 KASSERT(tso->in_len > 0, ("TSO input length went negative")); 886 KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); 887 888 n = min(tso->in_len, tso->packet_space); 889 890 tso->packet_space -= n; 891 tso->out_len -= n; 892 tso->in_len -= n; 893 894 desc = &txq->pend_desc[txq->n_pend_desc++]; 895 desc->eb_addr = tso->dma_addr; 896 desc->eb_size = n; 897 desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0; 898 899 tso->dma_addr += n; 900} 901 902/* Callback from bus_dmamap_load() for long TSO headers. */ 903static void tso_map_long_header(void *dma_addr_ret, 904 bus_dma_segment_t *segs, int nseg, 905 int error) 906{ 907 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && 908 __predict_true(nseg == 1)) ? 909 segs->ds_addr : 0); 910} 911 912/* 913 * tso_start_new_packet - generate a new header and prepare for the new packet 914 * 915 * Generate a new header and prepare for the new packet. Return 0 on 916 * success, or an error code if failed to alloc header. 917 */ 918static int tso_start_new_packet(struct sfxge_txq *txq, 919 struct sfxge_tso_state *tso, 920 unsigned int id) 921{ 922 struct sfxge_tx_mapping *stmp = &txq->stmp[id]; 923 struct tcphdr *tsoh_th; 924 unsigned ip_length; 925 caddr_t header; 926 uint64_t dma_addr; 927 bus_dmamap_t map; 928 efx_buffer_t *desc; 929 int rc; 930 931 /* Allocate a DMA-mapped header buffer. */ 932 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { 933 unsigned int page_index = (id / 2) / TSOH_PER_PAGE; 934 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; 935 936 header = (txq->tsoh_buffer[page_index].esm_base + 937 buf_index * TSOH_STD_SIZE); 938 dma_addr = (txq->tsoh_buffer[page_index].esm_addr + 939 buf_index * TSOH_STD_SIZE); 940 map = txq->tsoh_buffer[page_index].esm_map; 941 942 stmp->flags = 0; 943 } else { 944 /* We cannot use bus_dmamem_alloc() as that may sleep */ 945 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); 946 if (__predict_false(!header)) 947 return (ENOMEM); 948 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, 949 header, tso->header_len, 950 tso_map_long_header, &dma_addr, 951 BUS_DMA_NOWAIT); 952 if (__predict_false(dma_addr == 0)) { 953 if (rc == 0) { 954 /* Succeeded but got >1 segment */ 955 bus_dmamap_unload(txq->packet_dma_tag, 956 stmp->map); 957 rc = EINVAL; 958 } 959 free(header, M_SFXGE); 960 return (rc); 961 } 962 map = stmp->map; 963 964 txq->tso_long_headers++; 965 stmp->u.heap_buf = header; 966 stmp->flags = TX_BUF_UNMAP; 967 } 968 969 tsoh_th = (struct tcphdr *)(header + tso->tcph_off); 970 971 /* Copy and update the headers. */ 972 m_copydata(tso->mbuf, 0, tso->header_len, header); 973 974 tsoh_th->th_seq = htonl(tso->seqnum); 975 tso->seqnum += tso->seg_size; 976 if (tso->out_len > tso->seg_size) { 977 /* This packet will not finish the TSO burst. */ 978 ip_length = tso->header_len - tso->nh_off + tso->seg_size; 979 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); 980 } else { 981 /* This packet will be the last in the TSO burst. */ 982 ip_length = tso->header_len - tso->nh_off + tso->out_len; 983 } 984 985 if (tso->protocol == htons(ETHERTYPE_IP)) { 986 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); 987 tsoh_iph->ip_len = htons(ip_length); 988 /* XXX We should increment ip_id, but FreeBSD doesn't 989 * currently allocate extra IDs for multiple segments. 990 */ 991 } else { 992 struct ip6_hdr *tsoh_iph = 993 (struct ip6_hdr *)(header + tso->nh_off); 994 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); 995 } 996 997 /* Make the header visible to the hardware. */ 998 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); 999 1000 tso->packet_space = tso->seg_size; 1001 txq->tso_packets++; 1002 1003 /* Form a descriptor for this header. */ 1004 desc = &txq->pend_desc[txq->n_pend_desc++]; 1005 desc->eb_addr = dma_addr; 1006 desc->eb_size = tso->header_len; 1007 desc->eb_eop = 0; 1008 1009 return (0); 1010} 1011 1012static int 1013sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, 1014 const bus_dma_segment_t *dma_seg, int n_dma_seg) 1015{ 1016 struct sfxge_tso_state tso; 1017 unsigned int id, next_id; 1018 unsigned skipped = 0; 1019 1020 tso_start(&tso, mbuf); 1021 1022 while (dma_seg->ds_len + skipped <= tso.header_len) { 1023 skipped += dma_seg->ds_len; 1024 --n_dma_seg; 1025 KASSERT(n_dma_seg, ("no payload found in TSO packet")); 1026 ++dma_seg; 1027 } 1028 tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); 1029 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); 1030 1031 id = txq->added & txq->ptr_mask; 1032 if (__predict_false(tso_start_new_packet(txq, &tso, id))) 1033 return (-1); 1034 1035 while (1) { 1036 id = (id + 1) & txq->ptr_mask; 1037 tso_fill_packet_with_fragment(txq, &tso); 1038 1039 /* Move onto the next fragment? */ 1040 if (tso.in_len == 0) { 1041 --n_dma_seg; 1042 if (n_dma_seg == 0) 1043 break; 1044 ++dma_seg; 1045 tso.in_len = dma_seg->ds_len; 1046 tso.dma_addr = dma_seg->ds_addr; 1047 } 1048 1049 /* End of packet? */ 1050 if (tso.packet_space == 0) { 1051 /* If the queue is now full due to tiny MSS, 1052 * or we can't create another header, discard 1053 * the remainder of the input mbuf but do not 1054 * roll back the work we have done. 1055 */ 1056 if (txq->n_pend_desc + 1 /* header */ + n_dma_seg > 1057 SFXGE_TSO_MAX_DESC) { 1058 txq->tso_pdrop_too_many++; 1059 break; 1060 } 1061 next_id = (id + 1) & txq->ptr_mask; 1062 if (__predict_false(tso_start_new_packet(txq, &tso, 1063 next_id))) { 1064 txq->tso_pdrop_no_rsrc++; 1065 break; 1066 } 1067 id = next_id; 1068 } 1069 } 1070 1071 txq->tso_bursts++; 1072 return (id); 1073} 1074 1075static void 1076sfxge_tx_qunblock(struct sfxge_txq *txq) 1077{ 1078 struct sfxge_softc *sc; 1079 struct sfxge_evq *evq; 1080 1081 sc = txq->sc; 1082 evq = sc->evq[txq->evq_index]; 1083 1084 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); 1085 1086 if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) 1087 return; 1088 1089 SFXGE_TXQ_LOCK(txq); 1090 1091 if (txq->blocked) { 1092 unsigned int level; 1093 1094 level = txq->added - txq->completed; 1095 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) { 1096 /* reaped must be in sync with blocked */ 1097 sfxge_tx_qreap(txq); 1098 txq->blocked = 0; 1099 } 1100 } 1101 1102 sfxge_tx_qdpl_service(txq); 1103 /* note: lock has been dropped */ 1104} 1105 1106void 1107sfxge_tx_qflush_done(struct sfxge_txq *txq) 1108{ 1109 1110 txq->flush_state = SFXGE_FLUSH_DONE; 1111} 1112 1113static void 1114sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) 1115{ 1116 struct sfxge_txq *txq; 1117 struct sfxge_evq *evq; 1118 unsigned int count; 1119 1120 txq = sc->txq[index]; 1121 evq = sc->evq[txq->evq_index]; 1122 1123 SFXGE_TXQ_LOCK(txq); 1124 1125 KASSERT(txq->init_state == SFXGE_TXQ_STARTED, 1126 ("txq->init_state != SFXGE_TXQ_STARTED")); 1127 1128 txq->init_state = SFXGE_TXQ_INITIALIZED; 1129 txq->flush_state = SFXGE_FLUSH_PENDING; 1130 1131 /* Flush the transmit queue. */ 1132 efx_tx_qflush(txq->common); 1133 1134 SFXGE_TXQ_UNLOCK(txq); 1135 1136 count = 0; 1137 do { 1138 /* Spin for 100ms. */ 1139 DELAY(100000); 1140 1141 if (txq->flush_state != SFXGE_FLUSH_PENDING) 1142 break; 1143 } while (++count < 20); 1144 1145 SFXGE_EVQ_LOCK(evq); 1146 SFXGE_TXQ_LOCK(txq); 1147 1148 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, 1149 ("txq->flush_state == SFXGE_FLUSH_FAILED")); 1150 1151 txq->flush_state = SFXGE_FLUSH_DONE; 1152 1153 txq->blocked = 0; 1154 txq->pending = txq->added; 1155 1156 sfxge_tx_qcomplete(txq, evq); 1157 KASSERT(txq->completed == txq->added, 1158 ("txq->completed != txq->added")); 1159 1160 sfxge_tx_qreap(txq); 1161 KASSERT(txq->reaped == txq->completed, 1162 ("txq->reaped != txq->completed")); 1163 1164 txq->added = 0; 1165 txq->pending = 0; 1166 txq->completed = 0; 1167 txq->reaped = 0; 1168 1169 /* Destroy the common code transmit queue. */ 1170 efx_tx_qdestroy(txq->common); 1171 txq->common = NULL; 1172 1173 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1174 EFX_TXQ_NBUFS(sc->txq_entries)); 1175 1176 SFXGE_EVQ_UNLOCK(evq); 1177 SFXGE_TXQ_UNLOCK(txq); 1178} 1179 1180static int 1181sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) 1182{ 1183 struct sfxge_txq *txq; 1184 efsys_mem_t *esmp; 1185 uint16_t flags; 1186 struct sfxge_evq *evq; 1187 int rc; 1188 1189 txq = sc->txq[index]; 1190 esmp = &txq->mem; 1191 evq = sc->evq[txq->evq_index]; 1192 1193 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1194 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1195 KASSERT(evq->init_state == SFXGE_EVQ_STARTED, 1196 ("evq->init_state != SFXGE_EVQ_STARTED")); 1197 1198 /* Program the buffer table. */ 1199 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, 1200 EFX_TXQ_NBUFS(sc->txq_entries))) != 0) 1201 return (rc); 1202 1203 /* Determine the kind of queue we are creating. */ 1204 switch (txq->type) { 1205 case SFXGE_TXQ_NON_CKSUM: 1206 flags = 0; 1207 break; 1208 case SFXGE_TXQ_IP_CKSUM: 1209 flags = EFX_CKSUM_IPV4; 1210 break; 1211 case SFXGE_TXQ_IP_TCP_UDP_CKSUM: 1212 flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP; 1213 break; 1214 default: 1215 KASSERT(0, ("Impossible TX queue")); 1216 flags = 0; 1217 break; 1218 } 1219 1220 /* Create the common code transmit queue. */ 1221 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, 1222 sc->txq_entries, txq->buf_base_id, flags, evq->common, 1223 &txq->common)) != 0) 1224 goto fail; 1225 1226 SFXGE_TXQ_LOCK(txq); 1227 1228 /* Enable the transmit queue. */ 1229 efx_tx_qenable(txq->common); 1230 1231 txq->init_state = SFXGE_TXQ_STARTED; 1232 1233 SFXGE_TXQ_UNLOCK(txq); 1234 1235 return (0); 1236 1237fail: 1238 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, 1239 EFX_TXQ_NBUFS(sc->txq_entries)); 1240 return (rc); 1241} 1242 1243void 1244sfxge_tx_stop(struct sfxge_softc *sc) 1245{ 1246 int index; 1247 1248 index = sc->txq_count; 1249 while (--index >= 0) 1250 sfxge_tx_qstop(sc, index); 1251 1252 /* Tear down the transmit module */ 1253 efx_tx_fini(sc->enp); 1254} 1255 1256int 1257sfxge_tx_start(struct sfxge_softc *sc) 1258{ 1259 int index; 1260 int rc; 1261 1262 /* Initialize the common code transmit module. */ 1263 if ((rc = efx_tx_init(sc->enp)) != 0) 1264 return (rc); 1265 1266 for (index = 0; index < sc->txq_count; index++) { 1267 if ((rc = sfxge_tx_qstart(sc, index)) != 0) 1268 goto fail; 1269 } 1270 1271 return (0); 1272 1273fail: 1274 while (--index >= 0) 1275 sfxge_tx_qstop(sc, index); 1276 1277 efx_tx_fini(sc->enp); 1278 1279 return (rc); 1280} 1281 1282static int 1283sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node) 1284{ 1285 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev); 1286 struct sysctl_oid *stat_node; 1287 unsigned int id; 1288 1289 stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1290 "stats", CTLFLAG_RD, NULL, 1291 "Tx queue statistics"); 1292 if (stat_node == NULL) 1293 return (ENOMEM); 1294 1295 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1296 SYSCTL_ADD_ULONG( 1297 ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO, 1298 sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS, 1299 (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset), 1300 ""); 1301 } 1302 1303 return (0); 1304} 1305 1306/** 1307 * Destroy a transmit queue. 1308 */ 1309static void 1310sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) 1311{ 1312 struct sfxge_txq *txq; 1313 unsigned int nmaps; 1314 1315 txq = sc->txq[index]; 1316 1317 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, 1318 ("txq->init_state != SFXGE_TXQ_INITIALIZED")); 1319 1320 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) 1321 tso_fini(txq); 1322 1323 /* Free the context arrays. */ 1324 free(txq->pend_desc, M_SFXGE); 1325 nmaps = sc->txq_entries; 1326 while (nmaps-- != 0) 1327 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1328 free(txq->stmp, M_SFXGE); 1329 1330 /* Release DMA memory mapping. */ 1331 sfxge_dma_free(&txq->mem); 1332 1333 sc->txq[index] = NULL; 1334 1335 SFXGE_TXQ_LOCK_DESTROY(txq); 1336 1337 free(txq, M_SFXGE); 1338} 1339 1340static int 1341sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, 1342 enum sfxge_txq_type type, unsigned int evq_index) 1343{ 1344 char name[16]; 1345 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1346 struct sysctl_oid *txq_node; 1347 struct sfxge_txq *txq; 1348 struct sfxge_evq *evq; 1349 struct sfxge_tx_dpl *stdp; 1350 struct sysctl_oid *dpl_node; 1351 efsys_mem_t *esmp; 1352 unsigned int nmaps; 1353 int rc; 1354 1355 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); 1356 txq->sc = sc; 1357 txq->entries = sc->txq_entries; 1358 txq->ptr_mask = txq->entries - 1; 1359 1360 sc->txq[txq_index] = txq; 1361 esmp = &txq->mem; 1362 1363 evq = sc->evq[evq_index]; 1364 1365 /* Allocate and zero DMA space for the descriptor ring. */ 1366 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) 1367 return (rc); 1368 1369 /* Allocate buffer table entries. */ 1370 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), 1371 &txq->buf_base_id); 1372 1373 /* Create a DMA tag for packet mappings. */ 1374 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000, 1375 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, 1376 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL, 1377 &txq->packet_dma_tag) != 0) { 1378 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); 1379 rc = ENOMEM; 1380 goto fail; 1381 } 1382 1383 /* Allocate pending descriptor array for batching writes. */ 1384 txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries, 1385 M_SFXGE, M_ZERO | M_WAITOK); 1386 1387 /* Allocate and initialise mbuf DMA mapping array. */ 1388 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, 1389 M_SFXGE, M_ZERO | M_WAITOK); 1390 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { 1391 rc = bus_dmamap_create(txq->packet_dma_tag, 0, 1392 &txq->stmp[nmaps].map); 1393 if (rc != 0) 1394 goto fail2; 1395 } 1396 1397 snprintf(name, sizeof(name), "%u", txq_index); 1398 txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node), 1399 OID_AUTO, name, CTLFLAG_RD, NULL, ""); 1400 if (txq_node == NULL) { 1401 rc = ENOMEM; 1402 goto fail_txq_node; 1403 } 1404 1405 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && 1406 (rc = tso_init(txq)) != 0) 1407 goto fail3; 1408 1409 if (sfxge_tx_dpl_get_max <= 0) { 1410 log(LOG_ERR, "%s=%d must be greater than 0", 1411 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); 1412 rc = EINVAL; 1413 goto fail_tx_dpl_get_max; 1414 } 1415 if (sfxge_tx_dpl_get_non_tcp_max <= 0) { 1416 log(LOG_ERR, "%s=%d must be greater than 0", 1417 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, 1418 sfxge_tx_dpl_get_non_tcp_max); 1419 rc = EINVAL; 1420 goto fail_tx_dpl_get_max; 1421 } 1422 if (sfxge_tx_dpl_put_max < 0) { 1423 log(LOG_ERR, "%s=%d must be greater or equal to 0", 1424 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); 1425 rc = EINVAL; 1426 goto fail_tx_dpl_put_max; 1427 } 1428 1429 /* Initialize the deferred packet list. */ 1430 stdp = &txq->dpl; 1431 stdp->std_put_max = sfxge_tx_dpl_put_max; 1432 stdp->std_get_max = sfxge_tx_dpl_get_max; 1433 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; 1434 stdp->std_getp = &stdp->std_get; 1435 1436 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index); 1437 1438 dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, 1439 "dpl", CTLFLAG_RD, NULL, 1440 "Deferred packet list statistics"); 1441 if (dpl_node == NULL) { 1442 rc = ENOMEM; 1443 goto fail_dpl_node; 1444 } 1445 1446 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1447 "get_count", CTLFLAG_RD | CTLFLAG_STATS, 1448 &stdp->std_get_count, 0, ""); 1449 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1450 "get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, 1451 &stdp->std_get_non_tcp_count, 0, ""); 1452 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1453 "get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1454 &stdp->std_get_hiwat, 0, ""); 1455 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, 1456 "put_hiwat", CTLFLAG_RD | CTLFLAG_STATS, 1457 &stdp->std_put_hiwat, 0, ""); 1458 1459 rc = sfxge_txq_stat_init(txq, txq_node); 1460 if (rc != 0) 1461 goto fail_txq_stat_init; 1462 1463 txq->type = type; 1464 txq->evq_index = evq_index; 1465 txq->txq_index = txq_index; 1466 txq->init_state = SFXGE_TXQ_INITIALIZED; 1467 1468 return (0); 1469 1470fail_txq_stat_init: 1471fail_dpl_node: 1472fail_tx_dpl_put_max: 1473fail_tx_dpl_get_max: 1474fail3: 1475fail_txq_node: 1476 free(txq->pend_desc, M_SFXGE); 1477fail2: 1478 while (nmaps-- != 0) 1479 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); 1480 free(txq->stmp, M_SFXGE); 1481 bus_dma_tag_destroy(txq->packet_dma_tag); 1482 1483fail: 1484 sfxge_dma_free(esmp); 1485 1486 return (rc); 1487} 1488 1489static int 1490sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS) 1491{ 1492 struct sfxge_softc *sc = arg1; 1493 unsigned int id = arg2; 1494 unsigned long sum; 1495 unsigned int index; 1496 1497 /* Sum across all TX queues */ 1498 sum = 0; 1499 for (index = 0; index < sc->txq_count; index++) 1500 sum += *(unsigned long *)((caddr_t)sc->txq[index] + 1501 sfxge_tx_stats[id].offset); 1502 1503 return (SYSCTL_OUT(req, &sum, sizeof(sum))); 1504} 1505 1506static void 1507sfxge_tx_stat_init(struct sfxge_softc *sc) 1508{ 1509 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); 1510 struct sysctl_oid_list *stat_list; 1511 unsigned int id; 1512 1513 stat_list = SYSCTL_CHILDREN(sc->stats_node); 1514 1515 for (id = 0; id < nitems(sfxge_tx_stats); id++) { 1516 SYSCTL_ADD_PROC( 1517 ctx, stat_list, 1518 OID_AUTO, sfxge_tx_stats[id].name, 1519 CTLTYPE_ULONG|CTLFLAG_RD, 1520 sc, id, sfxge_tx_stat_handler, "LU", 1521 ""); 1522 } 1523} 1524 1525void 1526sfxge_tx_fini(struct sfxge_softc *sc) 1527{ 1528 int index; 1529 1530 index = sc->txq_count; 1531 while (--index >= 0) 1532 sfxge_tx_qfini(sc, index); 1533 1534 sc->txq_count = 0; 1535} 1536 1537 1538int 1539sfxge_tx_init(struct sfxge_softc *sc) 1540{ 1541 struct sfxge_intr *intr; 1542 int index; 1543 int rc; 1544 1545 intr = &sc->intr; 1546 1547 KASSERT(intr->state == SFXGE_INTR_INITIALIZED, 1548 ("intr->state != SFXGE_INTR_INITIALIZED")); 1549 1550 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc; 1551 1552 sc->txqs_node = SYSCTL_ADD_NODE( 1553 device_get_sysctl_ctx(sc->dev), 1554 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), 1555 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); 1556 if (sc->txqs_node == NULL) { 1557 rc = ENOMEM; 1558 goto fail_txq_node; 1559 } 1560 1561 /* Initialize the transmit queues */ 1562 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, 1563 SFXGE_TXQ_NON_CKSUM, 0)) != 0) 1564 goto fail; 1565 1566 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, 1567 SFXGE_TXQ_IP_CKSUM, 0)) != 0) 1568 goto fail2; 1569 1570 for (index = 0; 1571 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1; 1572 index++) { 1573 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index, 1574 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) 1575 goto fail3; 1576 } 1577 1578 sfxge_tx_stat_init(sc); 1579 1580 return (0); 1581 1582fail3: 1583 while (--index >= 0) 1584 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); 1585 1586 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); 1587 1588fail2: 1589 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); 1590 1591fail: 1592fail_txq_node: 1593 sc->txq_count = 0; 1594 return (rc); 1595} 1596