nicvf_queues.c revision 302408
1/* 2 * Copyright (C) 2015 Cavium Inc. 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 * $FreeBSD: stable/11/sys/dev/vnic/nicvf_queues.c 299446 2016-05-11 13:38:29Z zbb $ 27 * 28 */ 29#include <sys/cdefs.h> 30__FBSDID("$FreeBSD: stable/11/sys/dev/vnic/nicvf_queues.c 299446 2016-05-11 13:38:29Z zbb $"); 31 32#include "opt_inet.h" 33#include "opt_inet6.h" 34 35#include <sys/param.h> 36#include <sys/systm.h> 37#include <sys/bitset.h> 38#include <sys/bitstring.h> 39#include <sys/buf_ring.h> 40#include <sys/bus.h> 41#include <sys/endian.h> 42#include <sys/kernel.h> 43#include <sys/malloc.h> 44#include <sys/module.h> 45#include <sys/rman.h> 46#include <sys/pciio.h> 47#include <sys/pcpu.h> 48#include <sys/proc.h> 49#include <sys/sockio.h> 50#include <sys/socket.h> 51#include <sys/stdatomic.h> 52#include <sys/cpuset.h> 53#include <sys/lock.h> 54#include <sys/mutex.h> 55#include <sys/smp.h> 56#include <sys/taskqueue.h> 57 58#include <vm/vm.h> 59#include <vm/pmap.h> 60 61#include <machine/bus.h> 62#include <machine/vmparam.h> 63 64#include <net/if.h> 65#include <net/if_var.h> 66#include <net/if_media.h> 67#include <net/ifq.h> 68#include <net/bpf.h> 69#include <net/ethernet.h> 70 71#include <netinet/in_systm.h> 72#include <netinet/in.h> 73#include <netinet/if_ether.h> 74#include <netinet/ip.h> 75#include <netinet/ip6.h> 76#include <netinet/sctp.h> 77#include <netinet/tcp.h> 78#include <netinet/tcp_lro.h> 79#include <netinet/udp.h> 80 81#include <dev/pci/pcireg.h> 82#include <dev/pci/pcivar.h> 83 84#include "thunder_bgx.h" 85#include "nic_reg.h" 86#include "nic.h" 87#include "q_struct.h" 88#include "nicvf_queues.h" 89 90#define DEBUG 91#undef DEBUG 92 93#ifdef DEBUG 94#define dprintf(dev, fmt, ...) device_printf(dev, fmt, ##__VA_ARGS__) 95#else 96#define dprintf(dev, fmt, ...) 97#endif 98 99MALLOC_DECLARE(M_NICVF); 100 101static void nicvf_free_snd_queue(struct nicvf *, struct snd_queue *); 102static struct mbuf * nicvf_get_rcv_mbuf(struct nicvf *, struct cqe_rx_t *); 103static void nicvf_sq_disable(struct nicvf *, int); 104static void nicvf_sq_enable(struct nicvf *, struct snd_queue *, int); 105static void nicvf_put_sq_desc(struct snd_queue *, int); 106static void nicvf_cmp_queue_config(struct nicvf *, struct queue_set *, int, 107 boolean_t); 108static void nicvf_sq_free_used_descs(struct nicvf *, struct snd_queue *, int); 109 110static int nicvf_tx_mbuf_locked(struct snd_queue *, struct mbuf **); 111 112static void nicvf_rbdr_task(void *, int); 113static void nicvf_rbdr_task_nowait(void *, int); 114 115struct rbuf_info { 116 bus_dma_tag_t dmat; 117 bus_dmamap_t dmap; 118 struct mbuf * mbuf; 119}; 120 121#define GET_RBUF_INFO(x) ((struct rbuf_info *)((x) - NICVF_RCV_BUF_ALIGN_BYTES)) 122 123/* Poll a register for a specific value */ 124static int nicvf_poll_reg(struct nicvf *nic, int qidx, 125 uint64_t reg, int bit_pos, int bits, int val) 126{ 127 uint64_t bit_mask; 128 uint64_t reg_val; 129 int timeout = 10; 130 131 bit_mask = (1UL << bits) - 1; 132 bit_mask = (bit_mask << bit_pos); 133 134 while (timeout) { 135 reg_val = nicvf_queue_reg_read(nic, reg, qidx); 136 if (((reg_val & bit_mask) >> bit_pos) == val) 137 return (0); 138 139 DELAY(1000); 140 timeout--; 141 } 142 device_printf(nic->dev, "Poll on reg 0x%lx failed\n", reg); 143 return (ETIMEDOUT); 144} 145 146/* Callback for bus_dmamap_load() */ 147static void 148nicvf_dmamap_q_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) 149{ 150 bus_addr_t *paddr; 151 152 KASSERT(nseg == 1, ("wrong number of segments, should be 1")); 153 paddr = arg; 154 *paddr = segs->ds_addr; 155} 156 157/* Allocate memory for a queue's descriptors */ 158static int 159nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem, 160 int q_len, int desc_size, int align_bytes) 161{ 162 int err, err_dmat; 163 164 /* Create DMA tag first */ 165 err = bus_dma_tag_create( 166 bus_get_dma_tag(nic->dev), /* parent tag */ 167 align_bytes, /* alignment */ 168 0, /* boundary */ 169 BUS_SPACE_MAXADDR, /* lowaddr */ 170 BUS_SPACE_MAXADDR, /* highaddr */ 171 NULL, NULL, /* filtfunc, filtfuncarg */ 172 (q_len * desc_size), /* maxsize */ 173 1, /* nsegments */ 174 (q_len * desc_size), /* maxsegsize */ 175 0, /* flags */ 176 NULL, NULL, /* lockfunc, lockfuncarg */ 177 &dmem->dmat); /* dmat */ 178 179 if (err != 0) { 180 device_printf(nic->dev, 181 "Failed to create busdma tag for descriptors ring\n"); 182 return (err); 183 } 184 185 /* Allocate segment of continuous DMA safe memory */ 186 err = bus_dmamem_alloc( 187 dmem->dmat, /* DMA tag */ 188 &dmem->base, /* virtual address */ 189 (BUS_DMA_NOWAIT | BUS_DMA_ZERO), /* flags */ 190 &dmem->dmap); /* DMA map */ 191 if (err != 0) { 192 device_printf(nic->dev, "Failed to allocate DMA safe memory for" 193 "descriptors ring\n"); 194 goto dmamem_fail; 195 } 196 197 err = bus_dmamap_load( 198 dmem->dmat, 199 dmem->dmap, 200 dmem->base, 201 (q_len * desc_size), /* allocation size */ 202 nicvf_dmamap_q_cb, /* map to DMA address cb. */ 203 &dmem->phys_base, /* physical address */ 204 BUS_DMA_NOWAIT); 205 if (err != 0) { 206 device_printf(nic->dev, 207 "Cannot load DMA map of descriptors ring\n"); 208 goto dmamap_fail; 209 } 210 211 dmem->q_len = q_len; 212 dmem->size = (desc_size * q_len); 213 214 return (0); 215 216dmamap_fail: 217 bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap); 218 dmem->phys_base = 0; 219dmamem_fail: 220 err_dmat = bus_dma_tag_destroy(dmem->dmat); 221 dmem->base = NULL; 222 KASSERT(err_dmat == 0, 223 ("%s: Trying to destroy BUSY DMA tag", __func__)); 224 225 return (err); 226} 227 228/* Free queue's descriptor memory */ 229static void 230nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem) 231{ 232 int err; 233 234 if ((dmem == NULL) || (dmem->base == NULL)) 235 return; 236 237 /* Unload a map */ 238 bus_dmamap_sync(dmem->dmat, dmem->dmap, BUS_DMASYNC_POSTREAD); 239 bus_dmamap_unload(dmem->dmat, dmem->dmap); 240 /* Free DMA memory */ 241 bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap); 242 /* Destroy DMA tag */ 243 err = bus_dma_tag_destroy(dmem->dmat); 244 245 KASSERT(err == 0, 246 ("%s: Trying to destroy BUSY DMA tag", __func__)); 247 248 dmem->phys_base = 0; 249 dmem->base = NULL; 250} 251 252/* 253 * Allocate buffer for packet reception 254 * HW returns memory address where packet is DMA'ed but not a pointer 255 * into RBDR ring, so save buffer address at the start of fragment and 256 * align the start address to a cache aligned address 257 */ 258static __inline int 259nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr, 260 bus_dmamap_t dmap, int mflags, uint32_t buf_len, bus_addr_t *rbuf) 261{ 262 struct mbuf *mbuf; 263 struct rbuf_info *rinfo; 264 bus_dma_segment_t segs[1]; 265 int nsegs; 266 int err; 267 268 mbuf = m_getjcl(mflags, MT_DATA, M_PKTHDR, MCLBYTES); 269 if (mbuf == NULL) 270 return (ENOMEM); 271 272 /* 273 * The length is equal to the actual length + one 128b line 274 * used as a room for rbuf_info structure. 275 */ 276 mbuf->m_len = mbuf->m_pkthdr.len = buf_len; 277 278 err = bus_dmamap_load_mbuf_sg(rbdr->rbdr_buff_dmat, dmap, mbuf, segs, 279 &nsegs, BUS_DMA_NOWAIT); 280 if (err != 0) { 281 device_printf(nic->dev, 282 "Failed to map mbuf into DMA visible memory, err: %d\n", 283 err); 284 m_freem(mbuf); 285 bus_dmamap_destroy(rbdr->rbdr_buff_dmat, dmap); 286 return (err); 287 } 288 if (nsegs != 1) 289 panic("Unexpected number of DMA segments for RB: %d", nsegs); 290 /* 291 * Now use the room for rbuf_info structure 292 * and adjust mbuf data and length. 293 */ 294 rinfo = (struct rbuf_info *)mbuf->m_data; 295 m_adj(mbuf, NICVF_RCV_BUF_ALIGN_BYTES); 296 297 rinfo->dmat = rbdr->rbdr_buff_dmat; 298 rinfo->dmap = dmap; 299 rinfo->mbuf = mbuf; 300 301 *rbuf = segs[0].ds_addr + NICVF_RCV_BUF_ALIGN_BYTES; 302 303 return (0); 304} 305 306/* Retrieve mbuf for received packet */ 307static struct mbuf * 308nicvf_rb_ptr_to_mbuf(struct nicvf *nic, bus_addr_t rb_ptr) 309{ 310 struct mbuf *mbuf; 311 struct rbuf_info *rinfo; 312 313 /* Get buffer start address and alignment offset */ 314 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(rb_ptr)); 315 316 /* Now retrieve mbuf to give to stack */ 317 mbuf = rinfo->mbuf; 318 if (__predict_false(mbuf == NULL)) { 319 panic("%s: Received packet fragment with NULL mbuf", 320 device_get_nameunit(nic->dev)); 321 } 322 /* 323 * Clear the mbuf in the descriptor to indicate 324 * that this slot is processed and free to use. 325 */ 326 rinfo->mbuf = NULL; 327 328 bus_dmamap_sync(rinfo->dmat, rinfo->dmap, BUS_DMASYNC_POSTREAD); 329 bus_dmamap_unload(rinfo->dmat, rinfo->dmap); 330 331 return (mbuf); 332} 333 334/* Allocate RBDR ring and populate receive buffers */ 335static int 336nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr, int ring_len, 337 int buf_size, int qidx) 338{ 339 bus_dmamap_t dmap; 340 bus_addr_t rbuf; 341 struct rbdr_entry_t *desc; 342 int idx; 343 int err; 344 345 /* Allocate rbdr descriptors ring */ 346 err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len, 347 sizeof(struct rbdr_entry_t), NICVF_RCV_BUF_ALIGN_BYTES); 348 if (err != 0) { 349 device_printf(nic->dev, 350 "Failed to create RBDR descriptors ring\n"); 351 return (err); 352 } 353 354 rbdr->desc = rbdr->dmem.base; 355 /* 356 * Buffer size has to be in multiples of 128 bytes. 357 * Make room for metadata of size of one line (128 bytes). 358 */ 359 rbdr->dma_size = buf_size - NICVF_RCV_BUF_ALIGN_BYTES; 360 rbdr->enable = TRUE; 361 rbdr->thresh = RBDR_THRESH; 362 rbdr->nic = nic; 363 rbdr->idx = qidx; 364 365 /* 366 * Create DMA tag for Rx buffers. 367 * Each map created using this tag is intended to store Rx payload for 368 * one fragment and one header structure containing rbuf_info (thus 369 * additional 128 byte line since RB must be a multiple of 128 byte 370 * cache line). 371 */ 372 if (buf_size > MCLBYTES) { 373 device_printf(nic->dev, 374 "Buffer size to large for mbuf cluster\n"); 375 return (EINVAL); 376 } 377 err = bus_dma_tag_create( 378 bus_get_dma_tag(nic->dev), /* parent tag */ 379 NICVF_RCV_BUF_ALIGN_BYTES, /* alignment */ 380 0, /* boundary */ 381 DMAP_MAX_PHYSADDR, /* lowaddr */ 382 DMAP_MIN_PHYSADDR, /* highaddr */ 383 NULL, NULL, /* filtfunc, filtfuncarg */ 384 roundup2(buf_size, MCLBYTES), /* maxsize */ 385 1, /* nsegments */ 386 roundup2(buf_size, MCLBYTES), /* maxsegsize */ 387 0, /* flags */ 388 NULL, NULL, /* lockfunc, lockfuncarg */ 389 &rbdr->rbdr_buff_dmat); /* dmat */ 390 391 if (err != 0) { 392 device_printf(nic->dev, 393 "Failed to create busdma tag for RBDR buffers\n"); 394 return (err); 395 } 396 397 rbdr->rbdr_buff_dmaps = malloc(sizeof(*rbdr->rbdr_buff_dmaps) * 398 ring_len, M_NICVF, (M_WAITOK | M_ZERO)); 399 400 for (idx = 0; idx < ring_len; idx++) { 401 err = bus_dmamap_create(rbdr->rbdr_buff_dmat, 0, &dmap); 402 if (err != 0) { 403 device_printf(nic->dev, 404 "Failed to create DMA map for RB\n"); 405 return (err); 406 } 407 rbdr->rbdr_buff_dmaps[idx] = dmap; 408 409 err = nicvf_alloc_rcv_buffer(nic, rbdr, dmap, M_WAITOK, 410 DMA_BUFFER_LEN, &rbuf); 411 if (err != 0) 412 return (err); 413 414 desc = GET_RBDR_DESC(rbdr, idx); 415 desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN); 416 } 417 418 /* Allocate taskqueue */ 419 TASK_INIT(&rbdr->rbdr_task, 0, nicvf_rbdr_task, rbdr); 420 TASK_INIT(&rbdr->rbdr_task_nowait, 0, nicvf_rbdr_task_nowait, rbdr); 421 rbdr->rbdr_taskq = taskqueue_create_fast("nicvf_rbdr_taskq", M_WAITOK, 422 taskqueue_thread_enqueue, &rbdr->rbdr_taskq); 423 taskqueue_start_threads(&rbdr->rbdr_taskq, 1, PI_NET, "%s: rbdr_taskq", 424 device_get_nameunit(nic->dev)); 425 426 return (0); 427} 428 429/* Free RBDR ring and its receive buffers */ 430static void 431nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr) 432{ 433 struct mbuf *mbuf; 434 struct queue_set *qs; 435 struct rbdr_entry_t *desc; 436 struct rbuf_info *rinfo; 437 bus_addr_t buf_addr; 438 int head, tail, idx; 439 int err; 440 441 qs = nic->qs; 442 443 if ((qs == NULL) || (rbdr == NULL)) 444 return; 445 446 rbdr->enable = FALSE; 447 if (rbdr->rbdr_taskq != NULL) { 448 /* Remove tasks */ 449 while (taskqueue_cancel(rbdr->rbdr_taskq, 450 &rbdr->rbdr_task_nowait, NULL) != 0) { 451 /* Finish the nowait task first */ 452 taskqueue_drain(rbdr->rbdr_taskq, 453 &rbdr->rbdr_task_nowait); 454 } 455 taskqueue_free(rbdr->rbdr_taskq); 456 rbdr->rbdr_taskq = NULL; 457 458 while (taskqueue_cancel(taskqueue_thread, 459 &rbdr->rbdr_task, NULL) != 0) { 460 /* Now finish the sleepable task */ 461 taskqueue_drain(taskqueue_thread, &rbdr->rbdr_task); 462 } 463 } 464 465 /* 466 * Free all of the memory under the RB descriptors. 467 * There are assumptions here: 468 * 1. Corresponding RBDR is disabled 469 * - it is safe to operate using head and tail indexes 470 * 2. All bffers that were received are properly freed by 471 * the receive handler 472 * - there is no need to unload DMA map and free MBUF for other 473 * descriptors than unused ones 474 */ 475 if (rbdr->rbdr_buff_dmat != NULL) { 476 head = rbdr->head; 477 tail = rbdr->tail; 478 while (head != tail) { 479 desc = GET_RBDR_DESC(rbdr, head); 480 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN; 481 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr)); 482 bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap); 483 mbuf = rinfo->mbuf; 484 /* This will destroy everything including rinfo! */ 485 m_freem(mbuf); 486 head++; 487 head &= (rbdr->dmem.q_len - 1); 488 } 489 /* Free tail descriptor */ 490 desc = GET_RBDR_DESC(rbdr, tail); 491 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN; 492 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr)); 493 bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap); 494 mbuf = rinfo->mbuf; 495 /* This will destroy everything including rinfo! */ 496 m_freem(mbuf); 497 498 /* Destroy DMA maps */ 499 for (idx = 0; idx < qs->rbdr_len; idx++) { 500 if (rbdr->rbdr_buff_dmaps[idx] == NULL) 501 continue; 502 err = bus_dmamap_destroy(rbdr->rbdr_buff_dmat, 503 rbdr->rbdr_buff_dmaps[idx]); 504 KASSERT(err == 0, 505 ("%s: Could not destroy DMA map for RB, desc: %d", 506 __func__, idx)); 507 rbdr->rbdr_buff_dmaps[idx] = NULL; 508 } 509 510 /* Now destroy the tag */ 511 err = bus_dma_tag_destroy(rbdr->rbdr_buff_dmat); 512 KASSERT(err == 0, 513 ("%s: Trying to destroy BUSY DMA tag", __func__)); 514 515 rbdr->head = 0; 516 rbdr->tail = 0; 517 } 518 519 /* Free RBDR ring */ 520 nicvf_free_q_desc_mem(nic, &rbdr->dmem); 521} 522 523/* 524 * Refill receive buffer descriptors with new buffers. 525 */ 526static int 527nicvf_refill_rbdr(struct rbdr *rbdr, int mflags) 528{ 529 struct nicvf *nic; 530 struct queue_set *qs; 531 int rbdr_idx; 532 int tail, qcount; 533 int refill_rb_cnt; 534 struct rbdr_entry_t *desc; 535 bus_dmamap_t dmap; 536 bus_addr_t rbuf; 537 boolean_t rb_alloc_fail; 538 int new_rb; 539 540 rb_alloc_fail = TRUE; 541 new_rb = 0; 542 nic = rbdr->nic; 543 qs = nic->qs; 544 rbdr_idx = rbdr->idx; 545 546 /* Check if it's enabled */ 547 if (!rbdr->enable) 548 return (0); 549 550 /* Get no of desc's to be refilled */ 551 qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx); 552 qcount &= 0x7FFFF; 553 /* Doorbell can be ringed with a max of ring size minus 1 */ 554 if (qcount >= (qs->rbdr_len - 1)) { 555 rb_alloc_fail = FALSE; 556 goto out; 557 } else 558 refill_rb_cnt = qs->rbdr_len - qcount - 1; 559 560 /* Start filling descs from tail */ 561 tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3; 562 while (refill_rb_cnt) { 563 tail++; 564 tail &= (rbdr->dmem.q_len - 1); 565 566 dmap = rbdr->rbdr_buff_dmaps[tail]; 567 if (nicvf_alloc_rcv_buffer(nic, rbdr, dmap, mflags, 568 DMA_BUFFER_LEN, &rbuf)) { 569 /* Something went wrong. Resign */ 570 break; 571 } 572 desc = GET_RBDR_DESC(rbdr, tail); 573 desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN); 574 refill_rb_cnt--; 575 new_rb++; 576 } 577 578 /* make sure all memory stores are done before ringing doorbell */ 579 wmb(); 580 581 /* Check if buffer allocation failed */ 582 if (refill_rb_cnt == 0) 583 rb_alloc_fail = FALSE; 584 585 /* Notify HW */ 586 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, 587 rbdr_idx, new_rb); 588out: 589 if (!rb_alloc_fail) { 590 /* 591 * Re-enable RBDR interrupts only 592 * if buffer allocation is success. 593 */ 594 nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx); 595 596 return (0); 597 } 598 599 return (ENOMEM); 600} 601 602/* Refill RBs even if sleep is needed to reclaim memory */ 603static void 604nicvf_rbdr_task(void *arg, int pending) 605{ 606 struct rbdr *rbdr; 607 int err; 608 609 rbdr = (struct rbdr *)arg; 610 611 err = nicvf_refill_rbdr(rbdr, M_WAITOK); 612 if (__predict_false(err != 0)) { 613 panic("%s: Failed to refill RBs even when sleep enabled", 614 __func__); 615 } 616} 617 618/* Refill RBs as soon as possible without waiting */ 619static void 620nicvf_rbdr_task_nowait(void *arg, int pending) 621{ 622 struct rbdr *rbdr; 623 int err; 624 625 rbdr = (struct rbdr *)arg; 626 627 err = nicvf_refill_rbdr(rbdr, M_NOWAIT); 628 if (err != 0) { 629 /* 630 * Schedule another, sleepable kernel thread 631 * that will for sure refill the buffers. 632 */ 633 taskqueue_enqueue(taskqueue_thread, &rbdr->rbdr_task); 634 } 635} 636 637static int 638nicvf_rcv_pkt_handler(struct nicvf *nic, struct cmp_queue *cq, 639 struct cqe_rx_t *cqe_rx, int cqe_type) 640{ 641 struct mbuf *mbuf; 642 struct rcv_queue *rq; 643 int rq_idx; 644 int err = 0; 645 646 rq_idx = cqe_rx->rq_idx; 647 rq = &nic->qs->rq[rq_idx]; 648 649 /* Check for errors */ 650 err = nicvf_check_cqe_rx_errs(nic, cq, cqe_rx); 651 if (err && !cqe_rx->rb_cnt) 652 return (0); 653 654 mbuf = nicvf_get_rcv_mbuf(nic, cqe_rx); 655 if (mbuf == NULL) { 656 dprintf(nic->dev, "Packet not received\n"); 657 return (0); 658 } 659 660 /* If error packet */ 661 if (err != 0) { 662 m_freem(mbuf); 663 return (0); 664 } 665 666 if (rq->lro_enabled && 667 ((cqe_rx->l3_type == L3TYPE_IPV4) && (cqe_rx->l4_type == L4TYPE_TCP)) && 668 (mbuf->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) == 669 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) { 670 /* 671 * At this point it is known that there are no errors in the 672 * packet. Attempt to LRO enqueue. Send to stack if no resources 673 * or enqueue error. 674 */ 675 if ((rq->lro.lro_cnt != 0) && 676 (tcp_lro_rx(&rq->lro, mbuf, 0) == 0)) 677 return (0); 678 } 679 /* 680 * Push this packet to the stack later to avoid 681 * unlocking completion task in the middle of work. 682 */ 683 err = buf_ring_enqueue(cq->rx_br, mbuf); 684 if (err != 0) { 685 /* 686 * Failed to enqueue this mbuf. 687 * We don't drop it, just schedule another task. 688 */ 689 return (err); 690 } 691 692 return (0); 693} 694 695static void 696nicvf_snd_pkt_handler(struct nicvf *nic, struct cmp_queue *cq, 697 struct cqe_send_t *cqe_tx, int cqe_type) 698{ 699 bus_dmamap_t dmap; 700 struct mbuf *mbuf; 701 struct snd_queue *sq; 702 struct sq_hdr_subdesc *hdr; 703 704 mbuf = NULL; 705 sq = &nic->qs->sq[cqe_tx->sq_idx]; 706 707 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, cqe_tx->sqe_ptr); 708 if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) 709 return; 710 711 dprintf(nic->dev, 712 "%s Qset #%d SQ #%d SQ ptr #%d subdesc count %d\n", 713 __func__, cqe_tx->sq_qs, cqe_tx->sq_idx, 714 cqe_tx->sqe_ptr, hdr->subdesc_cnt); 715 716 dmap = (bus_dmamap_t)sq->snd_buff[cqe_tx->sqe_ptr].dmap; 717 bus_dmamap_unload(sq->snd_buff_dmat, dmap); 718 719 mbuf = (struct mbuf *)sq->snd_buff[cqe_tx->sqe_ptr].mbuf; 720 if (mbuf != NULL) { 721 m_freem(mbuf); 722 sq->snd_buff[cqe_tx->sqe_ptr].mbuf = NULL; 723 nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1); 724 } 725 726 nicvf_check_cqe_tx_errs(nic, cq, cqe_tx); 727} 728 729static int 730nicvf_cq_intr_handler(struct nicvf *nic, uint8_t cq_idx) 731{ 732 struct mbuf *mbuf; 733 struct ifnet *ifp; 734 int processed_cqe, work_done = 0, tx_done = 0; 735 int cqe_count, cqe_head; 736 struct queue_set *qs = nic->qs; 737 struct cmp_queue *cq = &qs->cq[cq_idx]; 738 struct snd_queue *sq = &qs->sq[cq_idx]; 739 struct rcv_queue *rq; 740 struct cqe_rx_t *cq_desc; 741 struct lro_ctrl *lro; 742 int rq_idx; 743 int cmp_err; 744 745 NICVF_CMP_LOCK(cq); 746 cmp_err = 0; 747 processed_cqe = 0; 748 /* Get no of valid CQ entries to process */ 749 cqe_count = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS, cq_idx); 750 cqe_count &= CQ_CQE_COUNT; 751 if (cqe_count == 0) 752 goto out; 753 754 /* Get head of the valid CQ entries */ 755 cqe_head = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_HEAD, cq_idx) >> 9; 756 cqe_head &= 0xFFFF; 757 758 dprintf(nic->dev, "%s CQ%d cqe_count %d cqe_head %d\n", 759 __func__, cq_idx, cqe_count, cqe_head); 760 while (processed_cqe < cqe_count) { 761 /* Get the CQ descriptor */ 762 cq_desc = (struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head); 763 cqe_head++; 764 cqe_head &= (cq->dmem.q_len - 1); 765 /* Prefetch next CQ descriptor */ 766 __builtin_prefetch((struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head)); 767 768 dprintf(nic->dev, "CQ%d cq_desc->cqe_type %d\n", cq_idx, 769 cq_desc->cqe_type); 770 switch (cq_desc->cqe_type) { 771 case CQE_TYPE_RX: 772 cmp_err = nicvf_rcv_pkt_handler(nic, cq, cq_desc, 773 CQE_TYPE_RX); 774 if (__predict_false(cmp_err != 0)) { 775 /* 776 * Ups. Cannot finish now. 777 * Let's try again later. 778 */ 779 goto done; 780 } 781 work_done++; 782 break; 783 case CQE_TYPE_SEND: 784 nicvf_snd_pkt_handler(nic, cq, (void *)cq_desc, 785 CQE_TYPE_SEND); 786 tx_done++; 787 break; 788 case CQE_TYPE_INVALID: 789 case CQE_TYPE_RX_SPLIT: 790 case CQE_TYPE_RX_TCP: 791 case CQE_TYPE_SEND_PTP: 792 /* Ignore for now */ 793 break; 794 } 795 processed_cqe++; 796 } 797done: 798 dprintf(nic->dev, 799 "%s CQ%d processed_cqe %d work_done %d\n", 800 __func__, cq_idx, processed_cqe, work_done); 801 802 /* Ring doorbell to inform H/W to reuse processed CQEs */ 803 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_DOOR, cq_idx, processed_cqe); 804 805 if ((tx_done > 0) && 806 ((if_getdrvflags(nic->ifp) & IFF_DRV_RUNNING) != 0)) { 807 /* Reenable TXQ if its stopped earlier due to SQ full */ 808 if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 809 taskqueue_enqueue(sq->snd_taskq, &sq->snd_task); 810 } 811out: 812 /* 813 * Flush any outstanding LRO work 814 */ 815 rq_idx = cq_idx; 816 rq = &nic->qs->rq[rq_idx]; 817 lro = &rq->lro; 818 tcp_lro_flush_all(lro); 819 820 NICVF_CMP_UNLOCK(cq); 821 822 ifp = nic->ifp; 823 /* Push received MBUFs to the stack */ 824 while (!buf_ring_empty(cq->rx_br)) { 825 mbuf = buf_ring_dequeue_mc(cq->rx_br); 826 if (__predict_true(mbuf != NULL)) 827 (*ifp->if_input)(ifp, mbuf); 828 } 829 830 return (cmp_err); 831} 832 833/* 834 * Qset error interrupt handler 835 * 836 * As of now only CQ errors are handled 837 */ 838static void 839nicvf_qs_err_task(void *arg, int pending) 840{ 841 struct nicvf *nic; 842 struct queue_set *qs; 843 int qidx; 844 uint64_t status; 845 boolean_t enable = TRUE; 846 847 nic = (struct nicvf *)arg; 848 qs = nic->qs; 849 850 /* Deactivate network interface */ 851 if_setdrvflagbits(nic->ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 852 853 /* Check if it is CQ err */ 854 for (qidx = 0; qidx < qs->cq_cnt; qidx++) { 855 status = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS, 856 qidx); 857 if ((status & CQ_ERR_MASK) == 0) 858 continue; 859 /* Process already queued CQEs and reconfig CQ */ 860 nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx); 861 nicvf_sq_disable(nic, qidx); 862 (void)nicvf_cq_intr_handler(nic, qidx); 863 nicvf_cmp_queue_config(nic, qs, qidx, enable); 864 nicvf_sq_free_used_descs(nic, &qs->sq[qidx], qidx); 865 nicvf_sq_enable(nic, &qs->sq[qidx], qidx); 866 nicvf_enable_intr(nic, NICVF_INTR_CQ, qidx); 867 } 868 869 if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 870 /* Re-enable Qset error interrupt */ 871 nicvf_enable_intr(nic, NICVF_INTR_QS_ERR, 0); 872} 873 874static void 875nicvf_cmp_task(void *arg, int pending) 876{ 877 struct cmp_queue *cq; 878 struct nicvf *nic; 879 int cmp_err; 880 881 cq = (struct cmp_queue *)arg; 882 nic = cq->nic; 883 884 /* Handle CQ descriptors */ 885 cmp_err = nicvf_cq_intr_handler(nic, cq->idx); 886 if (__predict_false(cmp_err != 0)) { 887 /* 888 * Schedule another thread here since we did not 889 * process the entire CQ due to Tx or Rx CQ parse error. 890 */ 891 taskqueue_enqueue(cq->cmp_taskq, &cq->cmp_task); 892 893 } 894 895 nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx); 896 /* Reenable interrupt (previously disabled in nicvf_intr_handler() */ 897 nicvf_enable_intr(nic, NICVF_INTR_CQ, cq->idx); 898 899} 900 901/* Initialize completion queue */ 902static int 903nicvf_init_cmp_queue(struct nicvf *nic, struct cmp_queue *cq, int q_len, 904 int qidx) 905{ 906 int err; 907 908 /* Initizalize lock */ 909 snprintf(cq->mtx_name, sizeof(cq->mtx_name), "%s: CQ(%d) lock", 910 device_get_nameunit(nic->dev), qidx); 911 mtx_init(&cq->mtx, cq->mtx_name, NULL, MTX_DEF); 912 913 err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE, 914 NICVF_CQ_BASE_ALIGN_BYTES); 915 916 if (err != 0) { 917 device_printf(nic->dev, 918 "Could not allocate DMA memory for CQ\n"); 919 return (err); 920 } 921 922 cq->desc = cq->dmem.base; 923 cq->thresh = pass1_silicon(nic->dev) ? 0 : CMP_QUEUE_CQE_THRESH; 924 cq->nic = nic; 925 cq->idx = qidx; 926 nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1; 927 928 cq->rx_br = buf_ring_alloc(CMP_QUEUE_LEN * 8, M_DEVBUF, M_WAITOK, 929 &cq->mtx); 930 931 /* Allocate taskqueue */ 932 TASK_INIT(&cq->cmp_task, 0, nicvf_cmp_task, cq); 933 cq->cmp_taskq = taskqueue_create_fast("nicvf_cmp_taskq", M_WAITOK, 934 taskqueue_thread_enqueue, &cq->cmp_taskq); 935 taskqueue_start_threads(&cq->cmp_taskq, 1, PI_NET, "%s: cmp_taskq(%d)", 936 device_get_nameunit(nic->dev), qidx); 937 938 return (0); 939} 940 941static void 942nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq) 943{ 944 945 if (cq == NULL) 946 return; 947 /* 948 * The completion queue itself should be disabled by now 949 * (ref. nicvf_snd_queue_config()). 950 * Ensure that it is safe to disable it or panic. 951 */ 952 if (cq->enable) 953 panic("%s: Trying to free working CQ(%d)", __func__, cq->idx); 954 955 if (cq->cmp_taskq != NULL) { 956 /* Remove task */ 957 while (taskqueue_cancel(cq->cmp_taskq, &cq->cmp_task, NULL) != 0) 958 taskqueue_drain(cq->cmp_taskq, &cq->cmp_task); 959 960 taskqueue_free(cq->cmp_taskq); 961 cq->cmp_taskq = NULL; 962 } 963 /* 964 * Completion interrupt will possibly enable interrupts again 965 * so disable interrupting now after we finished processing 966 * completion task. It is safe to do so since the corresponding CQ 967 * was already disabled. 968 */ 969 nicvf_disable_intr(nic, NICVF_INTR_CQ, cq->idx); 970 nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx); 971 972 NICVF_CMP_LOCK(cq); 973 nicvf_free_q_desc_mem(nic, &cq->dmem); 974 drbr_free(cq->rx_br, M_DEVBUF); 975 NICVF_CMP_UNLOCK(cq); 976 mtx_destroy(&cq->mtx); 977 memset(cq->mtx_name, 0, sizeof(cq->mtx_name)); 978} 979 980int 981nicvf_xmit_locked(struct snd_queue *sq) 982{ 983 struct nicvf *nic; 984 struct ifnet *ifp; 985 struct mbuf *next; 986 int err; 987 988 NICVF_TX_LOCK_ASSERT(sq); 989 990 nic = sq->nic; 991 ifp = nic->ifp; 992 err = 0; 993 994 while ((next = drbr_peek(ifp, sq->br)) != NULL) { 995 err = nicvf_tx_mbuf_locked(sq, &next); 996 if (err != 0) { 997 if (next == NULL) 998 drbr_advance(ifp, sq->br); 999 else 1000 drbr_putback(ifp, sq->br, next); 1001 1002 break; 1003 } 1004 drbr_advance(ifp, sq->br); 1005 /* Send a copy of the frame to the BPF listener */ 1006 ETHER_BPF_MTAP(ifp, next); 1007 } 1008 return (err); 1009} 1010 1011static void 1012nicvf_snd_task(void *arg, int pending) 1013{ 1014 struct snd_queue *sq = (struct snd_queue *)arg; 1015 struct nicvf *nic; 1016 struct ifnet *ifp; 1017 int err; 1018 1019 nic = sq->nic; 1020 ifp = nic->ifp; 1021 1022 /* 1023 * Skip sending anything if the driver is not running, 1024 * SQ full or link is down. 1025 */ 1026 if (((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != 1027 IFF_DRV_RUNNING) || !nic->link_up) 1028 return; 1029 1030 NICVF_TX_LOCK(sq); 1031 err = nicvf_xmit_locked(sq); 1032 NICVF_TX_UNLOCK(sq); 1033 /* Try again */ 1034 if (err != 0) 1035 taskqueue_enqueue(sq->snd_taskq, &sq->snd_task); 1036} 1037 1038/* Initialize transmit queue */ 1039static int 1040nicvf_init_snd_queue(struct nicvf *nic, struct snd_queue *sq, int q_len, 1041 int qidx) 1042{ 1043 size_t i; 1044 int err; 1045 1046 /* Initizalize TX lock for this queue */ 1047 snprintf(sq->mtx_name, sizeof(sq->mtx_name), "%s: SQ(%d) lock", 1048 device_get_nameunit(nic->dev), qidx); 1049 mtx_init(&sq->mtx, sq->mtx_name, NULL, MTX_DEF); 1050 1051 NICVF_TX_LOCK(sq); 1052 /* Allocate buffer ring */ 1053 sq->br = buf_ring_alloc(q_len / MIN_SQ_DESC_PER_PKT_XMIT, M_DEVBUF, 1054 M_NOWAIT, &sq->mtx); 1055 if (sq->br == NULL) { 1056 device_printf(nic->dev, 1057 "ERROR: Could not set up buf ring for SQ(%d)\n", qidx); 1058 err = ENOMEM; 1059 goto error; 1060 } 1061 1062 /* Allocate DMA memory for Tx descriptors */ 1063 err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE, 1064 NICVF_SQ_BASE_ALIGN_BYTES); 1065 if (err != 0) { 1066 device_printf(nic->dev, 1067 "Could not allocate DMA memory for SQ\n"); 1068 goto error; 1069 } 1070 1071 sq->desc = sq->dmem.base; 1072 sq->head = sq->tail = 0; 1073 atomic_store_rel_int(&sq->free_cnt, q_len - 1); 1074 sq->thresh = SND_QUEUE_THRESH; 1075 sq->idx = qidx; 1076 sq->nic = nic; 1077 1078 /* 1079 * Allocate DMA maps for Tx buffers 1080 */ 1081 1082 /* Create DMA tag first */ 1083 err = bus_dma_tag_create( 1084 bus_get_dma_tag(nic->dev), /* parent tag */ 1085 1, /* alignment */ 1086 0, /* boundary */ 1087 BUS_SPACE_MAXADDR, /* lowaddr */ 1088 BUS_SPACE_MAXADDR, /* highaddr */ 1089 NULL, NULL, /* filtfunc, filtfuncarg */ 1090 NICVF_TSO_MAXSIZE, /* maxsize */ 1091 NICVF_TSO_NSEGS, /* nsegments */ 1092 MCLBYTES, /* maxsegsize */ 1093 0, /* flags */ 1094 NULL, NULL, /* lockfunc, lockfuncarg */ 1095 &sq->snd_buff_dmat); /* dmat */ 1096 1097 if (err != 0) { 1098 device_printf(nic->dev, 1099 "Failed to create busdma tag for Tx buffers\n"); 1100 goto error; 1101 } 1102 1103 /* Allocate send buffers array */ 1104 sq->snd_buff = malloc(sizeof(*sq->snd_buff) * q_len, M_NICVF, 1105 (M_NOWAIT | M_ZERO)); 1106 if (sq->snd_buff == NULL) { 1107 device_printf(nic->dev, 1108 "Could not allocate memory for Tx buffers array\n"); 1109 err = ENOMEM; 1110 goto error; 1111 } 1112 1113 /* Now populate maps */ 1114 for (i = 0; i < q_len; i++) { 1115 err = bus_dmamap_create(sq->snd_buff_dmat, 0, 1116 &sq->snd_buff[i].dmap); 1117 if (err != 0) { 1118 device_printf(nic->dev, 1119 "Failed to create DMA maps for Tx buffers\n"); 1120 goto error; 1121 } 1122 } 1123 NICVF_TX_UNLOCK(sq); 1124 1125 /* Allocate taskqueue */ 1126 TASK_INIT(&sq->snd_task, 0, nicvf_snd_task, sq); 1127 sq->snd_taskq = taskqueue_create_fast("nicvf_snd_taskq", M_WAITOK, 1128 taskqueue_thread_enqueue, &sq->snd_taskq); 1129 taskqueue_start_threads(&sq->snd_taskq, 1, PI_NET, "%s: snd_taskq(%d)", 1130 device_get_nameunit(nic->dev), qidx); 1131 1132 return (0); 1133error: 1134 NICVF_TX_UNLOCK(sq); 1135 return (err); 1136} 1137 1138static void 1139nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq) 1140{ 1141 struct queue_set *qs = nic->qs; 1142 size_t i; 1143 int err; 1144 1145 if (sq == NULL) 1146 return; 1147 1148 if (sq->snd_taskq != NULL) { 1149 /* Remove task */ 1150 while (taskqueue_cancel(sq->snd_taskq, &sq->snd_task, NULL) != 0) 1151 taskqueue_drain(sq->snd_taskq, &sq->snd_task); 1152 1153 taskqueue_free(sq->snd_taskq); 1154 sq->snd_taskq = NULL; 1155 } 1156 1157 NICVF_TX_LOCK(sq); 1158 if (sq->snd_buff_dmat != NULL) { 1159 if (sq->snd_buff != NULL) { 1160 for (i = 0; i < qs->sq_len; i++) { 1161 m_freem(sq->snd_buff[i].mbuf); 1162 sq->snd_buff[i].mbuf = NULL; 1163 1164 bus_dmamap_unload(sq->snd_buff_dmat, 1165 sq->snd_buff[i].dmap); 1166 err = bus_dmamap_destroy(sq->snd_buff_dmat, 1167 sq->snd_buff[i].dmap); 1168 /* 1169 * If bus_dmamap_destroy fails it can cause 1170 * random panic later if the tag is also 1171 * destroyed in the process. 1172 */ 1173 KASSERT(err == 0, 1174 ("%s: Could not destroy DMA map for SQ", 1175 __func__)); 1176 } 1177 } 1178 1179 free(sq->snd_buff, M_NICVF); 1180 1181 err = bus_dma_tag_destroy(sq->snd_buff_dmat); 1182 KASSERT(err == 0, 1183 ("%s: Trying to destroy BUSY DMA tag", __func__)); 1184 } 1185 1186 /* Free private driver ring for this send queue */ 1187 if (sq->br != NULL) 1188 drbr_free(sq->br, M_DEVBUF); 1189 1190 if (sq->dmem.base != NULL) 1191 nicvf_free_q_desc_mem(nic, &sq->dmem); 1192 1193 NICVF_TX_UNLOCK(sq); 1194 /* Destroy Tx lock */ 1195 mtx_destroy(&sq->mtx); 1196 memset(sq->mtx_name, 0, sizeof(sq->mtx_name)); 1197} 1198 1199static void 1200nicvf_reclaim_snd_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1201{ 1202 1203 /* Disable send queue */ 1204 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0); 1205 /* Check if SQ is stopped */ 1206 if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01)) 1207 return; 1208 /* Reset send queue */ 1209 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 1210} 1211 1212static void 1213nicvf_reclaim_rcv_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1214{ 1215 union nic_mbx mbx = {}; 1216 1217 /* Make sure all packets in the pipeline are written back into mem */ 1218 mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC; 1219 nicvf_send_msg_to_pf(nic, &mbx); 1220} 1221 1222static void 1223nicvf_reclaim_cmp_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1224{ 1225 1226 /* Disable timer threshold (doesn't get reset upon CQ reset */ 1227 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0); 1228 /* Disable completion queue */ 1229 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0); 1230 /* Reset completion queue */ 1231 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 1232} 1233 1234static void 1235nicvf_reclaim_rbdr(struct nicvf *nic, struct rbdr *rbdr, int qidx) 1236{ 1237 uint64_t tmp, fifo_state; 1238 int timeout = 10; 1239 1240 /* Save head and tail pointers for feeing up buffers */ 1241 rbdr->head = 1242 nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_HEAD, qidx) >> 3; 1243 rbdr->tail = 1244 nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, qidx) >> 3; 1245 1246 /* 1247 * If RBDR FIFO is in 'FAIL' state then do a reset first 1248 * before relaiming. 1249 */ 1250 fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx); 1251 if (((fifo_state >> 62) & 0x03) == 0x3) { 1252 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, 1253 qidx, NICVF_RBDR_RESET); 1254 } 1255 1256 /* Disable RBDR */ 1257 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0); 1258 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 1259 return; 1260 while (1) { 1261 tmp = nicvf_queue_reg_read(nic, 1262 NIC_QSET_RBDR_0_1_PREFETCH_STATUS, qidx); 1263 if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF)) 1264 break; 1265 1266 DELAY(1000); 1267 timeout--; 1268 if (!timeout) { 1269 device_printf(nic->dev, 1270 "Failed polling on prefetch status\n"); 1271 return; 1272 } 1273 } 1274 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 1275 NICVF_RBDR_RESET); 1276 1277 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02)) 1278 return; 1279 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00); 1280 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 1281 return; 1282} 1283 1284/* Configures receive queue */ 1285static void 1286nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs, 1287 int qidx, bool enable) 1288{ 1289 union nic_mbx mbx = {}; 1290 struct rcv_queue *rq; 1291 struct rq_cfg rq_cfg; 1292 struct ifnet *ifp; 1293 struct lro_ctrl *lro; 1294 1295 ifp = nic->ifp; 1296 1297 rq = &qs->rq[qidx]; 1298 rq->enable = enable; 1299 1300 lro = &rq->lro; 1301 1302 /* Disable receive queue */ 1303 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0); 1304 1305 if (!rq->enable) { 1306 nicvf_reclaim_rcv_queue(nic, qs, qidx); 1307 /* Free LRO memory */ 1308 tcp_lro_free(lro); 1309 rq->lro_enabled = FALSE; 1310 return; 1311 } 1312 1313 /* Configure LRO if enabled */ 1314 rq->lro_enabled = FALSE; 1315 if ((if_getcapenable(ifp) & IFCAP_LRO) != 0) { 1316 if (tcp_lro_init(lro) != 0) { 1317 device_printf(nic->dev, 1318 "Failed to initialize LRO for RXQ%d\n", qidx); 1319 } else { 1320 rq->lro_enabled = TRUE; 1321 lro->ifp = nic->ifp; 1322 } 1323 } 1324 1325 rq->cq_qs = qs->vnic_id; 1326 rq->cq_idx = qidx; 1327 rq->start_rbdr_qs = qs->vnic_id; 1328 rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1; 1329 rq->cont_rbdr_qs = qs->vnic_id; 1330 rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1; 1331 /* all writes of RBDR data to be loaded into L2 Cache as well*/ 1332 rq->caching = 1; 1333 1334 /* Send a mailbox msg to PF to config RQ */ 1335 mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG; 1336 mbx.rq.qs_num = qs->vnic_id; 1337 mbx.rq.rq_num = qidx; 1338 mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) | 1339 (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) | 1340 (rq->cont_qs_rbdr_idx << 8) | (rq->start_rbdr_qs << 1) | 1341 (rq->start_qs_rbdr_idx); 1342 nicvf_send_msg_to_pf(nic, &mbx); 1343 1344 mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG; 1345 mbx.rq.cfg = (1UL << 63) | (1UL << 62) | (qs->vnic_id << 0); 1346 nicvf_send_msg_to_pf(nic, &mbx); 1347 1348 /* 1349 * RQ drop config 1350 * Enable CQ drop to reserve sufficient CQEs for all tx packets 1351 */ 1352 mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG; 1353 mbx.rq.cfg = (1UL << 62) | (RQ_CQ_DROP << 8); 1354 nicvf_send_msg_to_pf(nic, &mbx); 1355 1356 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, 0x00); 1357 1358 /* Enable Receive queue */ 1359 rq_cfg.ena = 1; 1360 rq_cfg.tcp_ena = 0; 1361 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 1362 *(uint64_t *)&rq_cfg); 1363} 1364 1365/* Configures completion queue */ 1366static void 1367nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs, 1368 int qidx, boolean_t enable) 1369{ 1370 struct cmp_queue *cq; 1371 struct cq_cfg cq_cfg; 1372 1373 cq = &qs->cq[qidx]; 1374 cq->enable = enable; 1375 1376 if (!cq->enable) { 1377 nicvf_reclaim_cmp_queue(nic, qs, qidx); 1378 return; 1379 } 1380 1381 /* Reset completion queue */ 1382 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 1383 1384 /* Set completion queue base address */ 1385 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE, qidx, 1386 (uint64_t)(cq->dmem.phys_base)); 1387 1388 /* Enable Completion queue */ 1389 cq_cfg.ena = 1; 1390 cq_cfg.reset = 0; 1391 cq_cfg.caching = 0; 1392 cq_cfg.qsize = CMP_QSIZE; 1393 cq_cfg.avg_con = 0; 1394 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(uint64_t *)&cq_cfg); 1395 1396 /* Set threshold value for interrupt generation */ 1397 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh); 1398 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 1399 nic->cq_coalesce_usecs); 1400} 1401 1402/* Configures transmit queue */ 1403static void 1404nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs, int qidx, 1405 boolean_t enable) 1406{ 1407 union nic_mbx mbx = {}; 1408 struct snd_queue *sq; 1409 struct sq_cfg sq_cfg; 1410 1411 sq = &qs->sq[qidx]; 1412 sq->enable = enable; 1413 1414 if (!sq->enable) { 1415 nicvf_reclaim_snd_queue(nic, qs, qidx); 1416 return; 1417 } 1418 1419 /* Reset send queue */ 1420 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 1421 1422 sq->cq_qs = qs->vnic_id; 1423 sq->cq_idx = qidx; 1424 1425 /* Send a mailbox msg to PF to config SQ */ 1426 mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG; 1427 mbx.sq.qs_num = qs->vnic_id; 1428 mbx.sq.sq_num = qidx; 1429 mbx.sq.sqs_mode = nic->sqs_mode; 1430 mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx; 1431 nicvf_send_msg_to_pf(nic, &mbx); 1432 1433 /* Set queue base address */ 1434 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE, qidx, 1435 (uint64_t)(sq->dmem.phys_base)); 1436 1437 /* Enable send queue & set queue size */ 1438 sq_cfg.ena = 1; 1439 sq_cfg.reset = 0; 1440 sq_cfg.ldwb = 0; 1441 sq_cfg.qsize = SND_QSIZE; 1442 sq_cfg.tstmp_bgx_intf = 0; 1443 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(uint64_t *)&sq_cfg); 1444 1445 /* Set threshold value for interrupt generation */ 1446 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh); 1447} 1448 1449/* Configures receive buffer descriptor ring */ 1450static void 1451nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs, int qidx, 1452 boolean_t enable) 1453{ 1454 struct rbdr *rbdr; 1455 struct rbdr_cfg rbdr_cfg; 1456 1457 rbdr = &qs->rbdr[qidx]; 1458 nicvf_reclaim_rbdr(nic, rbdr, qidx); 1459 if (!enable) 1460 return; 1461 1462 /* Set descriptor base address */ 1463 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE, qidx, 1464 (uint64_t)(rbdr->dmem.phys_base)); 1465 1466 /* Enable RBDR & set queue size */ 1467 /* Buffer size should be in multiples of 128 bytes */ 1468 rbdr_cfg.ena = 1; 1469 rbdr_cfg.reset = 0; 1470 rbdr_cfg.ldwb = 0; 1471 rbdr_cfg.qsize = RBDR_SIZE; 1472 rbdr_cfg.avg_con = 0; 1473 rbdr_cfg.lines = rbdr->dma_size / 128; 1474 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 1475 *(uint64_t *)&rbdr_cfg); 1476 1477 /* Notify HW */ 1478 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, qidx, 1479 qs->rbdr_len - 1); 1480 1481 /* Set threshold value for interrupt generation */ 1482 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH, qidx, 1483 rbdr->thresh - 1); 1484} 1485 1486/* Requests PF to assign and enable Qset */ 1487void 1488nicvf_qset_config(struct nicvf *nic, boolean_t enable) 1489{ 1490 union nic_mbx mbx = {}; 1491 struct queue_set *qs; 1492 struct qs_cfg *qs_cfg; 1493 1494 qs = nic->qs; 1495 if (qs == NULL) { 1496 device_printf(nic->dev, 1497 "Qset is still not allocated, don't init queues\n"); 1498 return; 1499 } 1500 1501 qs->enable = enable; 1502 qs->vnic_id = nic->vf_id; 1503 1504 /* Send a mailbox msg to PF to config Qset */ 1505 mbx.qs.msg = NIC_MBOX_MSG_QS_CFG; 1506 mbx.qs.num = qs->vnic_id; 1507 1508 mbx.qs.cfg = 0; 1509 qs_cfg = (struct qs_cfg *)&mbx.qs.cfg; 1510 if (qs->enable) { 1511 qs_cfg->ena = 1; 1512 qs_cfg->vnic = qs->vnic_id; 1513 } 1514 nicvf_send_msg_to_pf(nic, &mbx); 1515} 1516 1517static void 1518nicvf_free_resources(struct nicvf *nic) 1519{ 1520 int qidx; 1521 struct queue_set *qs; 1522 1523 qs = nic->qs; 1524 /* 1525 * Remove QS error task first since it has to be dead 1526 * to safely free completion queue tasks. 1527 */ 1528 if (qs->qs_err_taskq != NULL) { 1529 /* Shut down QS error tasks */ 1530 while (taskqueue_cancel(qs->qs_err_taskq, 1531 &qs->qs_err_task, NULL) != 0) { 1532 taskqueue_drain(qs->qs_err_taskq, &qs->qs_err_task); 1533 1534 } 1535 taskqueue_free(qs->qs_err_taskq); 1536 qs->qs_err_taskq = NULL; 1537 } 1538 /* Free receive buffer descriptor ring */ 1539 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1540 nicvf_free_rbdr(nic, &qs->rbdr[qidx]); 1541 1542 /* Free completion queue */ 1543 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1544 nicvf_free_cmp_queue(nic, &qs->cq[qidx]); 1545 1546 /* Free send queue */ 1547 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1548 nicvf_free_snd_queue(nic, &qs->sq[qidx]); 1549} 1550 1551static int 1552nicvf_alloc_resources(struct nicvf *nic) 1553{ 1554 struct queue_set *qs = nic->qs; 1555 int qidx; 1556 1557 /* Alloc receive buffer descriptor ring */ 1558 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) { 1559 if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len, 1560 DMA_BUFFER_LEN, qidx)) 1561 goto alloc_fail; 1562 } 1563 1564 /* Alloc send queue */ 1565 for (qidx = 0; qidx < qs->sq_cnt; qidx++) { 1566 if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx)) 1567 goto alloc_fail; 1568 } 1569 1570 /* Alloc completion queue */ 1571 for (qidx = 0; qidx < qs->cq_cnt; qidx++) { 1572 if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len, qidx)) 1573 goto alloc_fail; 1574 } 1575 1576 /* Allocate QS error taskqueue */ 1577 TASK_INIT(&qs->qs_err_task, 0, nicvf_qs_err_task, nic); 1578 qs->qs_err_taskq = taskqueue_create_fast("nicvf_qs_err_taskq", M_WAITOK, 1579 taskqueue_thread_enqueue, &qs->qs_err_taskq); 1580 taskqueue_start_threads(&qs->qs_err_taskq, 1, PI_NET, "%s: qs_taskq", 1581 device_get_nameunit(nic->dev)); 1582 1583 return (0); 1584alloc_fail: 1585 nicvf_free_resources(nic); 1586 return (ENOMEM); 1587} 1588 1589int 1590nicvf_set_qset_resources(struct nicvf *nic) 1591{ 1592 struct queue_set *qs; 1593 1594 qs = malloc(sizeof(*qs), M_NICVF, (M_ZERO | M_WAITOK)); 1595 nic->qs = qs; 1596 1597 /* Set count of each queue */ 1598 qs->rbdr_cnt = RBDR_CNT; 1599 qs->rq_cnt = RCV_QUEUE_CNT; 1600 1601 qs->sq_cnt = SND_QUEUE_CNT; 1602 qs->cq_cnt = CMP_QUEUE_CNT; 1603 1604 /* Set queue lengths */ 1605 qs->rbdr_len = RCV_BUF_COUNT; 1606 qs->sq_len = SND_QUEUE_LEN; 1607 qs->cq_len = CMP_QUEUE_LEN; 1608 1609 nic->rx_queues = qs->rq_cnt; 1610 nic->tx_queues = qs->sq_cnt; 1611 1612 return (0); 1613} 1614 1615int 1616nicvf_config_data_transfer(struct nicvf *nic, boolean_t enable) 1617{ 1618 boolean_t disable = FALSE; 1619 struct queue_set *qs; 1620 int qidx; 1621 1622 qs = nic->qs; 1623 if (qs == NULL) 1624 return (0); 1625 1626 if (enable) { 1627 if (nicvf_alloc_resources(nic) != 0) 1628 return (ENOMEM); 1629 1630 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1631 nicvf_snd_queue_config(nic, qs, qidx, enable); 1632 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1633 nicvf_cmp_queue_config(nic, qs, qidx, enable); 1634 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1635 nicvf_rbdr_config(nic, qs, qidx, enable); 1636 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1637 nicvf_rcv_queue_config(nic, qs, qidx, enable); 1638 } else { 1639 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1640 nicvf_rcv_queue_config(nic, qs, qidx, disable); 1641 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1642 nicvf_rbdr_config(nic, qs, qidx, disable); 1643 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1644 nicvf_snd_queue_config(nic, qs, qidx, disable); 1645 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1646 nicvf_cmp_queue_config(nic, qs, qidx, disable); 1647 1648 nicvf_free_resources(nic); 1649 } 1650 1651 return (0); 1652} 1653 1654/* 1655 * Get a free desc from SQ 1656 * returns descriptor ponter & descriptor number 1657 */ 1658static __inline int 1659nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt) 1660{ 1661 int qentry; 1662 1663 qentry = sq->tail; 1664 atomic_subtract_int(&sq->free_cnt, desc_cnt); 1665 sq->tail += desc_cnt; 1666 sq->tail &= (sq->dmem.q_len - 1); 1667 1668 return (qentry); 1669} 1670 1671/* Free descriptor back to SQ for future use */ 1672static void 1673nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt) 1674{ 1675 1676 atomic_add_int(&sq->free_cnt, desc_cnt); 1677 sq->head += desc_cnt; 1678 sq->head &= (sq->dmem.q_len - 1); 1679} 1680 1681static __inline int 1682nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry) 1683{ 1684 qentry++; 1685 qentry &= (sq->dmem.q_len - 1); 1686 return (qentry); 1687} 1688 1689static void 1690nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx) 1691{ 1692 uint64_t sq_cfg; 1693 1694 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1695 sq_cfg |= NICVF_SQ_EN; 1696 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1697 /* Ring doorbell so that H/W restarts processing SQEs */ 1698 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0); 1699} 1700 1701static void 1702nicvf_sq_disable(struct nicvf *nic, int qidx) 1703{ 1704 uint64_t sq_cfg; 1705 1706 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1707 sq_cfg &= ~NICVF_SQ_EN; 1708 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1709} 1710 1711static void 1712nicvf_sq_free_used_descs(struct nicvf *nic, struct snd_queue *sq, int qidx) 1713{ 1714 uint64_t head, tail; 1715 struct snd_buff *snd_buff; 1716 struct sq_hdr_subdesc *hdr; 1717 1718 NICVF_TX_LOCK(sq); 1719 head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4; 1720 tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4; 1721 while (sq->head != head) { 1722 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head); 1723 if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) { 1724 nicvf_put_sq_desc(sq, 1); 1725 continue; 1726 } 1727 snd_buff = &sq->snd_buff[sq->head]; 1728 if (snd_buff->mbuf != NULL) { 1729 bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap); 1730 m_freem(snd_buff->mbuf); 1731 sq->snd_buff[sq->head].mbuf = NULL; 1732 } 1733 nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1); 1734 } 1735 NICVF_TX_UNLOCK(sq); 1736} 1737 1738/* 1739 * Add SQ HEADER subdescriptor. 1740 * First subdescriptor for every send descriptor. 1741 */ 1742static __inline int 1743nicvf_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry, 1744 int subdesc_cnt, struct mbuf *mbuf, int len) 1745{ 1746 struct nicvf *nic; 1747 struct sq_hdr_subdesc *hdr; 1748 struct ether_vlan_header *eh; 1749#ifdef INET 1750 struct ip *ip; 1751 struct tcphdr *th; 1752#endif 1753 uint16_t etype; 1754 int ehdrlen, iphlen, poff; 1755 1756 nic = sq->nic; 1757 1758 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry); 1759 sq->snd_buff[qentry].mbuf = mbuf; 1760 1761 memset(hdr, 0, SND_QUEUE_DESC_SIZE); 1762 hdr->subdesc_type = SQ_DESC_TYPE_HEADER; 1763 /* Enable notification via CQE after processing SQE */ 1764 hdr->post_cqe = 1; 1765 /* No of subdescriptors following this */ 1766 hdr->subdesc_cnt = subdesc_cnt; 1767 hdr->tot_len = len; 1768 1769 eh = mtod(mbuf, struct ether_vlan_header *); 1770 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { 1771 ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 1772 etype = ntohs(eh->evl_proto); 1773 } else { 1774 ehdrlen = ETHER_HDR_LEN; 1775 etype = ntohs(eh->evl_encap_proto); 1776 } 1777 1778 switch (etype) { 1779#ifdef INET6 1780 case ETHERTYPE_IPV6: 1781 /* ARM64TODO: Add support for IPv6 */ 1782 hdr->csum_l3 = 0; 1783 sq->snd_buff[qentry].mbuf = NULL; 1784 return (ENXIO); 1785#endif 1786#ifdef INET 1787 case ETHERTYPE_IP: 1788 if (mbuf->m_len < ehdrlen + sizeof(struct ip)) { 1789 mbuf = m_pullup(mbuf, ehdrlen + sizeof(struct ip)); 1790 sq->snd_buff[qentry].mbuf = mbuf; 1791 if (mbuf == NULL) 1792 return (ENOBUFS); 1793 } 1794 1795 ip = (struct ip *)(mbuf->m_data + ehdrlen); 1796 iphlen = ip->ip_hl << 2; 1797 poff = ehdrlen + iphlen; 1798 1799 if (mbuf->m_pkthdr.csum_flags != 0) { 1800 hdr->csum_l3 = 1; /* Enable IP csum calculation */ 1801 switch (ip->ip_p) { 1802 case IPPROTO_TCP: 1803 if ((mbuf->m_pkthdr.csum_flags & CSUM_TCP) == 0) 1804 break; 1805 1806 if (mbuf->m_len < (poff + sizeof(struct tcphdr))) { 1807 mbuf = m_pullup(mbuf, poff + sizeof(struct tcphdr)); 1808 sq->snd_buff[qentry].mbuf = mbuf; 1809 if (mbuf == NULL) 1810 return (ENOBUFS); 1811 } 1812 hdr->csum_l4 = SEND_L4_CSUM_TCP; 1813 break; 1814 case IPPROTO_UDP: 1815 if ((mbuf->m_pkthdr.csum_flags & CSUM_UDP) == 0) 1816 break; 1817 1818 if (mbuf->m_len < (poff + sizeof(struct udphdr))) { 1819 mbuf = m_pullup(mbuf, poff + sizeof(struct udphdr)); 1820 sq->snd_buff[qentry].mbuf = mbuf; 1821 if (mbuf == NULL) 1822 return (ENOBUFS); 1823 } 1824 hdr->csum_l4 = SEND_L4_CSUM_UDP; 1825 break; 1826 case IPPROTO_SCTP: 1827 if ((mbuf->m_pkthdr.csum_flags & CSUM_SCTP) == 0) 1828 break; 1829 1830 if (mbuf->m_len < (poff + sizeof(struct sctphdr))) { 1831 mbuf = m_pullup(mbuf, poff + sizeof(struct sctphdr)); 1832 sq->snd_buff[qentry].mbuf = mbuf; 1833 if (mbuf == NULL) 1834 return (ENOBUFS); 1835 } 1836 hdr->csum_l4 = SEND_L4_CSUM_SCTP; 1837 break; 1838 default: 1839 break; 1840 } 1841 hdr->l3_offset = ehdrlen; 1842 hdr->l4_offset = ehdrlen + iphlen; 1843 } 1844 1845 if ((mbuf->m_pkthdr.tso_segsz != 0) && nic->hw_tso) { 1846 /* 1847 * Extract ip again as m_data could have been modified. 1848 */ 1849 ip = (struct ip *)(mbuf->m_data + ehdrlen); 1850 th = (struct tcphdr *)((caddr_t)ip + iphlen); 1851 1852 hdr->tso = 1; 1853 hdr->tso_start = ehdrlen + iphlen + (th->th_off * 4); 1854 hdr->tso_max_paysize = mbuf->m_pkthdr.tso_segsz; 1855 hdr->inner_l3_offset = ehdrlen - 2; 1856 nic->drv_stats.tx_tso++; 1857 } 1858 break; 1859#endif 1860 default: 1861 hdr->csum_l3 = 0; 1862 } 1863 1864 return (0); 1865} 1866 1867/* 1868 * SQ GATHER subdescriptor 1869 * Must follow HDR descriptor 1870 */ 1871static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry, 1872 int size, uint64_t data) 1873{ 1874 struct sq_gather_subdesc *gather; 1875 1876 qentry &= (sq->dmem.q_len - 1); 1877 gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry); 1878 1879 memset(gather, 0, SND_QUEUE_DESC_SIZE); 1880 gather->subdesc_type = SQ_DESC_TYPE_GATHER; 1881 gather->ld_type = NIC_SEND_LD_TYPE_E_LDD; 1882 gather->size = size; 1883 gather->addr = data; 1884} 1885 1886/* Put an mbuf to a SQ for packet transfer. */ 1887static int 1888nicvf_tx_mbuf_locked(struct snd_queue *sq, struct mbuf **mbufp) 1889{ 1890 bus_dma_segment_t segs[256]; 1891 struct snd_buff *snd_buff; 1892 size_t seg; 1893 int nsegs, qentry; 1894 int subdesc_cnt; 1895 int err; 1896 1897 NICVF_TX_LOCK_ASSERT(sq); 1898 1899 if (sq->free_cnt == 0) 1900 return (ENOBUFS); 1901 1902 snd_buff = &sq->snd_buff[sq->tail]; 1903 1904 err = bus_dmamap_load_mbuf_sg(sq->snd_buff_dmat, snd_buff->dmap, 1905 *mbufp, segs, &nsegs, BUS_DMA_NOWAIT); 1906 if (__predict_false(err != 0)) { 1907 /* ARM64TODO: Add mbuf defragmenting if we lack maps */ 1908 m_freem(*mbufp); 1909 *mbufp = NULL; 1910 return (err); 1911 } 1912 1913 /* Set how many subdescriptors is required */ 1914 subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT + nsegs - 1; 1915 if (subdesc_cnt > sq->free_cnt) { 1916 /* ARM64TODO: Add mbuf defragmentation if we lack descriptors */ 1917 bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap); 1918 return (ENOBUFS); 1919 } 1920 1921 qentry = nicvf_get_sq_desc(sq, subdesc_cnt); 1922 1923 /* Add SQ header subdesc */ 1924 err = nicvf_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, *mbufp, 1925 (*mbufp)->m_pkthdr.len); 1926 if (err != 0) { 1927 nicvf_put_sq_desc(sq, subdesc_cnt); 1928 bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap); 1929 if (err == ENOBUFS) { 1930 m_freem(*mbufp); 1931 *mbufp = NULL; 1932 } 1933 return (err); 1934 } 1935 1936 /* Add SQ gather subdescs */ 1937 for (seg = 0; seg < nsegs; seg++) { 1938 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1939 nicvf_sq_add_gather_subdesc(sq, qentry, segs[seg].ds_len, 1940 segs[seg].ds_addr); 1941 } 1942 1943 /* make sure all memory stores are done before ringing doorbell */ 1944 bus_dmamap_sync(sq->dmem.dmat, sq->dmem.dmap, BUS_DMASYNC_PREWRITE); 1945 1946 dprintf(sq->nic->dev, "%s: sq->idx: %d, subdesc_cnt: %d\n", 1947 __func__, sq->idx, subdesc_cnt); 1948 /* Inform HW to xmit new packet */ 1949 nicvf_queue_reg_write(sq->nic, NIC_QSET_SQ_0_7_DOOR, 1950 sq->idx, subdesc_cnt); 1951 return (0); 1952} 1953 1954static __inline u_int 1955frag_num(u_int i) 1956{ 1957#if BYTE_ORDER == BIG_ENDIAN 1958 return ((i & ~3) + 3 - (i & 3)); 1959#else 1960 return (i); 1961#endif 1962} 1963 1964/* Returns MBUF for a received packet */ 1965struct mbuf * 1966nicvf_get_rcv_mbuf(struct nicvf *nic, struct cqe_rx_t *cqe_rx) 1967{ 1968 int frag; 1969 int payload_len = 0; 1970 struct mbuf *mbuf; 1971 struct mbuf *mbuf_frag; 1972 uint16_t *rb_lens = NULL; 1973 uint64_t *rb_ptrs = NULL; 1974 1975 mbuf = NULL; 1976 rb_lens = (uint16_t *)((uint8_t *)cqe_rx + (3 * sizeof(uint64_t))); 1977 rb_ptrs = (uint64_t *)((uint8_t *)cqe_rx + (6 * sizeof(uint64_t))); 1978 1979 dprintf(nic->dev, "%s rb_cnt %d rb0_ptr %lx rb0_sz %d\n", 1980 __func__, cqe_rx->rb_cnt, cqe_rx->rb0_ptr, cqe_rx->rb0_sz); 1981 1982 for (frag = 0; frag < cqe_rx->rb_cnt; frag++) { 1983 payload_len = rb_lens[frag_num(frag)]; 1984 if (frag == 0) { 1985 /* First fragment */ 1986 mbuf = nicvf_rb_ptr_to_mbuf(nic, 1987 (*rb_ptrs - cqe_rx->align_pad)); 1988 mbuf->m_len = payload_len; 1989 mbuf->m_data += cqe_rx->align_pad; 1990 if_setrcvif(mbuf, nic->ifp); 1991 } else { 1992 /* Add fragments */ 1993 mbuf_frag = nicvf_rb_ptr_to_mbuf(nic, *rb_ptrs); 1994 m_append(mbuf, payload_len, mbuf_frag->m_data); 1995 m_freem(mbuf_frag); 1996 } 1997 /* Next buffer pointer */ 1998 rb_ptrs++; 1999 } 2000 2001 if (__predict_true(mbuf != NULL)) { 2002 m_fixhdr(mbuf); 2003 mbuf->m_pkthdr.flowid = cqe_rx->rq_idx; 2004 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE); 2005 if (__predict_true((if_getcapenable(nic->ifp) & IFCAP_RXCSUM) != 0)) { 2006 /* 2007 * HW by default verifies IP & TCP/UDP/SCTP checksums 2008 */ 2009 if (__predict_true(cqe_rx->l3_type == L3TYPE_IPV4)) { 2010 mbuf->m_pkthdr.csum_flags = 2011 (CSUM_IP_CHECKED | CSUM_IP_VALID); 2012 } 2013 2014 switch (cqe_rx->l4_type) { 2015 case L4TYPE_UDP: 2016 case L4TYPE_TCP: /* fall through */ 2017 mbuf->m_pkthdr.csum_flags |= 2018 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 2019 mbuf->m_pkthdr.csum_data = 0xffff; 2020 break; 2021 case L4TYPE_SCTP: 2022 mbuf->m_pkthdr.csum_flags |= CSUM_SCTP_VALID; 2023 break; 2024 default: 2025 break; 2026 } 2027 } 2028 } 2029 2030 return (mbuf); 2031} 2032 2033/* Enable interrupt */ 2034void 2035nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx) 2036{ 2037 uint64_t reg_val; 2038 2039 reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S); 2040 2041 switch (int_type) { 2042 case NICVF_INTR_CQ: 2043 reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2044 break; 2045 case NICVF_INTR_SQ: 2046 reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2047 break; 2048 case NICVF_INTR_RBDR: 2049 reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2050 break; 2051 case NICVF_INTR_PKT_DROP: 2052 reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2053 break; 2054 case NICVF_INTR_TCP_TIMER: 2055 reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2056 break; 2057 case NICVF_INTR_MBOX: 2058 reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT); 2059 break; 2060 case NICVF_INTR_QS_ERR: 2061 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2062 break; 2063 default: 2064 device_printf(nic->dev, 2065 "Failed to enable interrupt: unknown type\n"); 2066 break; 2067 } 2068 2069 nicvf_reg_write(nic, NIC_VF_ENA_W1S, reg_val); 2070} 2071 2072/* Disable interrupt */ 2073void 2074nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx) 2075{ 2076 uint64_t reg_val = 0; 2077 2078 switch (int_type) { 2079 case NICVF_INTR_CQ: 2080 reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2081 break; 2082 case NICVF_INTR_SQ: 2083 reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2084 break; 2085 case NICVF_INTR_RBDR: 2086 reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2087 break; 2088 case NICVF_INTR_PKT_DROP: 2089 reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2090 break; 2091 case NICVF_INTR_TCP_TIMER: 2092 reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2093 break; 2094 case NICVF_INTR_MBOX: 2095 reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT); 2096 break; 2097 case NICVF_INTR_QS_ERR: 2098 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2099 break; 2100 default: 2101 device_printf(nic->dev, 2102 "Failed to disable interrupt: unknown type\n"); 2103 break; 2104 } 2105 2106 nicvf_reg_write(nic, NIC_VF_ENA_W1C, reg_val); 2107} 2108 2109/* Clear interrupt */ 2110void 2111nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx) 2112{ 2113 uint64_t reg_val = 0; 2114 2115 switch (int_type) { 2116 case NICVF_INTR_CQ: 2117 reg_val = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2118 break; 2119 case NICVF_INTR_SQ: 2120 reg_val = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2121 break; 2122 case NICVF_INTR_RBDR: 2123 reg_val = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2124 break; 2125 case NICVF_INTR_PKT_DROP: 2126 reg_val = (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2127 break; 2128 case NICVF_INTR_TCP_TIMER: 2129 reg_val = (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2130 break; 2131 case NICVF_INTR_MBOX: 2132 reg_val = (1UL << NICVF_INTR_MBOX_SHIFT); 2133 break; 2134 case NICVF_INTR_QS_ERR: 2135 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2136 break; 2137 default: 2138 device_printf(nic->dev, 2139 "Failed to clear interrupt: unknown type\n"); 2140 break; 2141 } 2142 2143 nicvf_reg_write(nic, NIC_VF_INT, reg_val); 2144} 2145 2146/* Check if interrupt is enabled */ 2147int 2148nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx) 2149{ 2150 uint64_t reg_val; 2151 uint64_t mask = 0xff; 2152 2153 reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S); 2154 2155 switch (int_type) { 2156 case NICVF_INTR_CQ: 2157 mask = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2158 break; 2159 case NICVF_INTR_SQ: 2160 mask = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2161 break; 2162 case NICVF_INTR_RBDR: 2163 mask = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2164 break; 2165 case NICVF_INTR_PKT_DROP: 2166 mask = NICVF_INTR_PKT_DROP_MASK; 2167 break; 2168 case NICVF_INTR_TCP_TIMER: 2169 mask = NICVF_INTR_TCP_TIMER_MASK; 2170 break; 2171 case NICVF_INTR_MBOX: 2172 mask = NICVF_INTR_MBOX_MASK; 2173 break; 2174 case NICVF_INTR_QS_ERR: 2175 mask = NICVF_INTR_QS_ERR_MASK; 2176 break; 2177 default: 2178 device_printf(nic->dev, 2179 "Failed to check interrupt enable: unknown type\n"); 2180 break; 2181 } 2182 2183 return (reg_val & mask); 2184} 2185 2186void 2187nicvf_update_rq_stats(struct nicvf *nic, int rq_idx) 2188{ 2189 struct rcv_queue *rq; 2190 2191#define GET_RQ_STATS(reg) \ 2192 nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\ 2193 (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 2194 2195 rq = &nic->qs->rq[rq_idx]; 2196 rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS); 2197 rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS); 2198} 2199 2200void 2201nicvf_update_sq_stats(struct nicvf *nic, int sq_idx) 2202{ 2203 struct snd_queue *sq; 2204 2205#define GET_SQ_STATS(reg) \ 2206 nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\ 2207 (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 2208 2209 sq = &nic->qs->sq[sq_idx]; 2210 sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS); 2211 sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS); 2212} 2213 2214/* Check for errors in the receive cmp.queue entry */ 2215int 2216nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cmp_queue *cq, 2217 struct cqe_rx_t *cqe_rx) 2218{ 2219 struct nicvf_hw_stats *stats = &nic->hw_stats; 2220 struct nicvf_drv_stats *drv_stats = &nic->drv_stats; 2221 2222 if (!cqe_rx->err_level && !cqe_rx->err_opcode) { 2223 drv_stats->rx_frames_ok++; 2224 return (0); 2225 } 2226 2227 switch (cqe_rx->err_opcode) { 2228 case CQ_RX_ERROP_RE_PARTIAL: 2229 stats->rx_bgx_truncated_pkts++; 2230 break; 2231 case CQ_RX_ERROP_RE_JABBER: 2232 stats->rx_jabber_errs++; 2233 break; 2234 case CQ_RX_ERROP_RE_FCS: 2235 stats->rx_fcs_errs++; 2236 break; 2237 case CQ_RX_ERROP_RE_RX_CTL: 2238 stats->rx_bgx_errs++; 2239 break; 2240 case CQ_RX_ERROP_PREL2_ERR: 2241 stats->rx_prel2_errs++; 2242 break; 2243 case CQ_RX_ERROP_L2_MAL: 2244 stats->rx_l2_hdr_malformed++; 2245 break; 2246 case CQ_RX_ERROP_L2_OVERSIZE: 2247 stats->rx_oversize++; 2248 break; 2249 case CQ_RX_ERROP_L2_UNDERSIZE: 2250 stats->rx_undersize++; 2251 break; 2252 case CQ_RX_ERROP_L2_LENMISM: 2253 stats->rx_l2_len_mismatch++; 2254 break; 2255 case CQ_RX_ERROP_L2_PCLP: 2256 stats->rx_l2_pclp++; 2257 break; 2258 case CQ_RX_ERROP_IP_NOT: 2259 stats->rx_ip_ver_errs++; 2260 break; 2261 case CQ_RX_ERROP_IP_CSUM_ERR: 2262 stats->rx_ip_csum_errs++; 2263 break; 2264 case CQ_RX_ERROP_IP_MAL: 2265 stats->rx_ip_hdr_malformed++; 2266 break; 2267 case CQ_RX_ERROP_IP_MALD: 2268 stats->rx_ip_payload_malformed++; 2269 break; 2270 case CQ_RX_ERROP_IP_HOP: 2271 stats->rx_ip_ttl_errs++; 2272 break; 2273 case CQ_RX_ERROP_L3_PCLP: 2274 stats->rx_l3_pclp++; 2275 break; 2276 case CQ_RX_ERROP_L4_MAL: 2277 stats->rx_l4_malformed++; 2278 break; 2279 case CQ_RX_ERROP_L4_CHK: 2280 stats->rx_l4_csum_errs++; 2281 break; 2282 case CQ_RX_ERROP_UDP_LEN: 2283 stats->rx_udp_len_errs++; 2284 break; 2285 case CQ_RX_ERROP_L4_PORT: 2286 stats->rx_l4_port_errs++; 2287 break; 2288 case CQ_RX_ERROP_TCP_FLAG: 2289 stats->rx_tcp_flag_errs++; 2290 break; 2291 case CQ_RX_ERROP_TCP_OFFSET: 2292 stats->rx_tcp_offset_errs++; 2293 break; 2294 case CQ_RX_ERROP_L4_PCLP: 2295 stats->rx_l4_pclp++; 2296 break; 2297 case CQ_RX_ERROP_RBDR_TRUNC: 2298 stats->rx_truncated_pkts++; 2299 break; 2300 } 2301 2302 return (1); 2303} 2304 2305/* Check for errors in the send cmp.queue entry */ 2306int 2307nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cmp_queue *cq, 2308 struct cqe_send_t *cqe_tx) 2309{ 2310 struct cmp_queue_stats *stats = &cq->stats; 2311 2312 switch (cqe_tx->send_status) { 2313 case CQ_TX_ERROP_GOOD: 2314 stats->tx.good++; 2315 return (0); 2316 case CQ_TX_ERROP_DESC_FAULT: 2317 stats->tx.desc_fault++; 2318 break; 2319 case CQ_TX_ERROP_HDR_CONS_ERR: 2320 stats->tx.hdr_cons_err++; 2321 break; 2322 case CQ_TX_ERROP_SUBDC_ERR: 2323 stats->tx.subdesc_err++; 2324 break; 2325 case CQ_TX_ERROP_IMM_SIZE_OFLOW: 2326 stats->tx.imm_size_oflow++; 2327 break; 2328 case CQ_TX_ERROP_DATA_SEQUENCE_ERR: 2329 stats->tx.data_seq_err++; 2330 break; 2331 case CQ_TX_ERROP_MEM_SEQUENCE_ERR: 2332 stats->tx.mem_seq_err++; 2333 break; 2334 case CQ_TX_ERROP_LOCK_VIOL: 2335 stats->tx.lock_viol++; 2336 break; 2337 case CQ_TX_ERROP_DATA_FAULT: 2338 stats->tx.data_fault++; 2339 break; 2340 case CQ_TX_ERROP_TSTMP_CONFLICT: 2341 stats->tx.tstmp_conflict++; 2342 break; 2343 case CQ_TX_ERROP_TSTMP_TIMEOUT: 2344 stats->tx.tstmp_timeout++; 2345 break; 2346 case CQ_TX_ERROP_MEM_FAULT: 2347 stats->tx.mem_fault++; 2348 break; 2349 case CQ_TX_ERROP_CK_OVERLAP: 2350 stats->tx.csum_overlap++; 2351 break; 2352 case CQ_TX_ERROP_CK_OFLOW: 2353 stats->tx.csum_overflow++; 2354 break; 2355 } 2356 2357 return (1); 2358} 2359