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 * $FreeBSD$ 27 * 28 */ 29 #include <sys/cdefs.h> 30 __FBSDID("$FreeBSD$"); 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 <netinet6/ip6_var.h> 82 83 #include <dev/pci/pcireg.h> 84 #include <dev/pci/pcivar.h> 85 86 #include "thunder_bgx.h" 87 #include "nic_reg.h" 88 #include "nic.h" 89 #include "q_struct.h" 90 #include "nicvf_queues.h" 91 92 #define DEBUG 93 #undef DEBUG 94 95 #ifdef DEBUG 96 #define dprintf(dev, fmt, ...) device_printf(dev, fmt, ##__VA_ARGS__) 97 #else 98 #define dprintf(dev, fmt, ...) 99 #endif 100 101 MALLOC_DECLARE(M_NICVF); 102 103 static void nicvf_free_snd_queue(struct nicvf *, struct snd_queue *); 104 static struct mbuf * nicvf_get_rcv_mbuf(struct nicvf *, struct cqe_rx_t *); 105 static void nicvf_sq_disable(struct nicvf *, int); 106 static void nicvf_sq_enable(struct nicvf *, struct snd_queue *, int); 107 static void nicvf_put_sq_desc(struct snd_queue *, int); 108 static void nicvf_cmp_queue_config(struct nicvf *, struct queue_set *, int, 109 boolean_t); 110 static void nicvf_sq_free_used_descs(struct nicvf *, struct snd_queue *, int); 111 112 static int nicvf_tx_mbuf_locked(struct snd_queue *, struct mbuf **); 113 114 static void nicvf_rbdr_task(void *, int); 115 static void nicvf_rbdr_task_nowait(void *, int); 116 117 struct rbuf_info { 118 bus_dma_tag_t dmat; 119 bus_dmamap_t dmap; 120 struct mbuf * mbuf; 121 }; 122 123 #define GET_RBUF_INFO(x) ((struct rbuf_info *)((x) - NICVF_RCV_BUF_ALIGN_BYTES)) 124 125 /* Poll a register for a specific value */ 126 static int nicvf_poll_reg(struct nicvf *nic, int qidx, 127 uint64_t reg, int bit_pos, int bits, int val) 128 { 129 uint64_t bit_mask; 130 uint64_t reg_val; 131 int timeout = 10; 132 133 bit_mask = (1UL << bits) - 1; 134 bit_mask = (bit_mask << bit_pos); 135 136 while (timeout) { 137 reg_val = nicvf_queue_reg_read(nic, reg, qidx); 138 if (((reg_val & bit_mask) >> bit_pos) == val) 139 return (0); 140 141 DELAY(1000); 142 timeout--; 143 } 144 device_printf(nic->dev, "Poll on reg 0x%lx failed\n", reg); 145 return (ETIMEDOUT); 146 } 147 148 /* Callback for bus_dmamap_load() */ 149 static void 150 nicvf_dmamap_q_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) 151 { 152 bus_addr_t *paddr; 153 154 KASSERT(nseg == 1, ("wrong number of segments, should be 1")); 155 paddr = arg; 156 *paddr = segs->ds_addr; 157 } 158 159 /* Allocate memory for a queue's descriptors */ 160 static int 161 nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem, 162 int q_len, int desc_size, int align_bytes) 163 { 164 int err, err_dmat; 165 166 /* Create DMA tag first */ 167 err = bus_dma_tag_create( 168 bus_get_dma_tag(nic->dev), /* parent tag */ 169 align_bytes, /* alignment */ 170 0, /* boundary */ 171 BUS_SPACE_MAXADDR, /* lowaddr */ 172 BUS_SPACE_MAXADDR, /* highaddr */ 173 NULL, NULL, /* filtfunc, filtfuncarg */ 174 (q_len * desc_size), /* maxsize */ 175 1, /* nsegments */ 176 (q_len * desc_size), /* maxsegsize */ 177 0, /* flags */ 178 NULL, NULL, /* lockfunc, lockfuncarg */ 179 &dmem->dmat); /* dmat */ 180 181 if (err != 0) { 182 device_printf(nic->dev, 183 "Failed to create busdma tag for descriptors ring\n"); 184 return (err); 185 } 186 187 /* Allocate segment of continuous DMA safe memory */ 188 err = bus_dmamem_alloc( 189 dmem->dmat, /* DMA tag */ 190 &dmem->base, /* virtual address */ 191 (BUS_DMA_NOWAIT | BUS_DMA_ZERO), /* flags */ 192 &dmem->dmap); /* DMA map */ 193 if (err != 0) { 194 device_printf(nic->dev, "Failed to allocate DMA safe memory for" 195 "descriptors ring\n"); 196 goto dmamem_fail; 197 } 198 199 err = bus_dmamap_load( 200 dmem->dmat, 201 dmem->dmap, 202 dmem->base, 203 (q_len * desc_size), /* allocation size */ 204 nicvf_dmamap_q_cb, /* map to DMA address cb. */ 205 &dmem->phys_base, /* physical address */ 206 BUS_DMA_NOWAIT); 207 if (err != 0) { 208 device_printf(nic->dev, 209 "Cannot load DMA map of descriptors ring\n"); 210 goto dmamap_fail; 211 } 212 213 dmem->q_len = q_len; 214 dmem->size = (desc_size * q_len); 215 216 return (0); 217 218 dmamap_fail: 219 bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap); 220 dmem->phys_base = 0; 221 dmamem_fail: 222 err_dmat = bus_dma_tag_destroy(dmem->dmat); 223 dmem->base = NULL; 224 KASSERT(err_dmat == 0, 225 ("%s: Trying to destroy BUSY DMA tag", __func__)); 226 227 return (err); 228 } 229 230 /* Free queue's descriptor memory */ 231 static void 232 nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem) 233 { 234 int err; 235 236 if ((dmem == NULL) || (dmem->base == NULL)) 237 return; 238 239 /* Unload a map */ 240 bus_dmamap_sync(dmem->dmat, dmem->dmap, BUS_DMASYNC_POSTREAD); 241 bus_dmamap_unload(dmem->dmat, dmem->dmap); 242 /* Free DMA memory */ 243 bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap); 244 /* Destroy DMA tag */ 245 err = bus_dma_tag_destroy(dmem->dmat); 246 247 KASSERT(err == 0, 248 ("%s: Trying to destroy BUSY DMA tag", __func__)); 249 250 dmem->phys_base = 0; 251 dmem->base = NULL; 252 } 253 254 /* 255 * Allocate buffer for packet reception 256 * HW returns memory address where packet is DMA'ed but not a pointer 257 * into RBDR ring, so save buffer address at the start of fragment and 258 * align the start address to a cache aligned address 259 */ 260 static __inline int 261 nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr, 262 bus_dmamap_t dmap, int mflags, uint32_t buf_len, bus_addr_t *rbuf) 263 { 264 struct mbuf *mbuf; 265 struct rbuf_info *rinfo; 266 bus_dma_segment_t segs[1]; 267 int nsegs; 268 int err; 269 270 mbuf = m_getjcl(mflags, MT_DATA, M_PKTHDR, MCLBYTES); 271 if (mbuf == NULL) 272 return (ENOMEM); 273 274 /* 275 * The length is equal to the actual length + one 128b line 276 * used as a room for rbuf_info structure. 277 */ 278 mbuf->m_len = mbuf->m_pkthdr.len = buf_len; 279 280 err = bus_dmamap_load_mbuf_sg(rbdr->rbdr_buff_dmat, dmap, mbuf, segs, 281 &nsegs, BUS_DMA_NOWAIT); 282 if (err != 0) { 283 device_printf(nic->dev, 284 "Failed to map mbuf into DMA visible memory, err: %d\n", 285 err); 286 m_freem(mbuf); 287 bus_dmamap_destroy(rbdr->rbdr_buff_dmat, dmap); 288 return (err); 289 } 290 if (nsegs != 1) 291 panic("Unexpected number of DMA segments for RB: %d", nsegs); 292 /* 293 * Now use the room for rbuf_info structure 294 * and adjust mbuf data and length. 295 */ 296 rinfo = (struct rbuf_info *)mbuf->m_data; 297 m_adj(mbuf, NICVF_RCV_BUF_ALIGN_BYTES); 298 299 rinfo->dmat = rbdr->rbdr_buff_dmat; 300 rinfo->dmap = dmap; 301 rinfo->mbuf = mbuf; 302 303 *rbuf = segs[0].ds_addr + NICVF_RCV_BUF_ALIGN_BYTES; 304 305 return (0); 306 } 307 308 /* Retrieve mbuf for received packet */ 309 static struct mbuf * 310 nicvf_rb_ptr_to_mbuf(struct nicvf *nic, bus_addr_t rb_ptr) 311 { 312 struct mbuf *mbuf; 313 struct rbuf_info *rinfo; 314 315 /* Get buffer start address and alignment offset */ 316 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(rb_ptr)); 317 318 /* Now retrieve mbuf to give to stack */ 319 mbuf = rinfo->mbuf; 320 if (__predict_false(mbuf == NULL)) { 321 panic("%s: Received packet fragment with NULL mbuf", 322 device_get_nameunit(nic->dev)); 323 } 324 /* 325 * Clear the mbuf in the descriptor to indicate 326 * that this slot is processed and free to use. 327 */ 328 rinfo->mbuf = NULL; 329 330 bus_dmamap_sync(rinfo->dmat, rinfo->dmap, BUS_DMASYNC_POSTREAD); 331 bus_dmamap_unload(rinfo->dmat, rinfo->dmap); 332 333 return (mbuf); 334 } 335 336 /* Allocate RBDR ring and populate receive buffers */ 337 static int 338 nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr, int ring_len, 339 int buf_size, int qidx) 340 { 341 bus_dmamap_t dmap; 342 bus_addr_t rbuf; 343 struct rbdr_entry_t *desc; 344 int idx; 345 int err; 346 347 /* Allocate rbdr descriptors ring */ 348 err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len, 349 sizeof(struct rbdr_entry_t), NICVF_RCV_BUF_ALIGN_BYTES); 350 if (err != 0) { 351 device_printf(nic->dev, 352 "Failed to create RBDR descriptors ring\n"); 353 return (err); 354 } 355 356 rbdr->desc = rbdr->dmem.base; 357 /* 358 * Buffer size has to be in multiples of 128 bytes. 359 * Make room for metadata of size of one line (128 bytes). 360 */ 361 rbdr->dma_size = buf_size - NICVF_RCV_BUF_ALIGN_BYTES; 362 rbdr->enable = TRUE; 363 rbdr->thresh = RBDR_THRESH; 364 rbdr->nic = nic; 365 rbdr->idx = qidx; 366 367 /* 368 * Create DMA tag for Rx buffers. 369 * Each map created using this tag is intended to store Rx payload for 370 * one fragment and one header structure containing rbuf_info (thus 371 * additional 128 byte line since RB must be a multiple of 128 byte 372 * cache line). 373 */ 374 if (buf_size > MCLBYTES) { 375 device_printf(nic->dev, 376 "Buffer size to large for mbuf cluster\n"); 377 return (EINVAL); 378 } 379 err = bus_dma_tag_create( 380 bus_get_dma_tag(nic->dev), /* parent tag */ 381 NICVF_RCV_BUF_ALIGN_BYTES, /* alignment */ 382 0, /* boundary */ 383 DMAP_MAX_PHYSADDR, /* lowaddr */ 384 DMAP_MIN_PHYSADDR, /* highaddr */ 385 NULL, NULL, /* filtfunc, filtfuncarg */ 386 roundup2(buf_size, MCLBYTES), /* maxsize */ 387 1, /* nsegments */ 388 roundup2(buf_size, MCLBYTES), /* maxsegsize */ 389 0, /* flags */ 390 NULL, NULL, /* lockfunc, lockfuncarg */ 391 &rbdr->rbdr_buff_dmat); /* dmat */ 392 393 if (err != 0) { 394 device_printf(nic->dev, 395 "Failed to create busdma tag for RBDR buffers\n"); 396 return (err); 397 } 398 399 rbdr->rbdr_buff_dmaps = malloc(sizeof(*rbdr->rbdr_buff_dmaps) * 400 ring_len, M_NICVF, (M_WAITOK | M_ZERO)); 401 402 for (idx = 0; idx < ring_len; idx++) { 403 err = bus_dmamap_create(rbdr->rbdr_buff_dmat, 0, &dmap); 404 if (err != 0) { 405 device_printf(nic->dev, 406 "Failed to create DMA map for RB\n"); 407 return (err); 408 } 409 rbdr->rbdr_buff_dmaps[idx] = dmap; 410 411 err = nicvf_alloc_rcv_buffer(nic, rbdr, dmap, M_WAITOK, 412 DMA_BUFFER_LEN, &rbuf); 413 if (err != 0) 414 return (err); 415 416 desc = GET_RBDR_DESC(rbdr, idx); 417 desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN); 418 } 419 420 /* Allocate taskqueue */ 421 TASK_INIT(&rbdr->rbdr_task, 0, nicvf_rbdr_task, rbdr); 422 TASK_INIT(&rbdr->rbdr_task_nowait, 0, nicvf_rbdr_task_nowait, rbdr); 423 rbdr->rbdr_taskq = taskqueue_create_fast("nicvf_rbdr_taskq", M_WAITOK, 424 taskqueue_thread_enqueue, &rbdr->rbdr_taskq); 425 taskqueue_start_threads(&rbdr->rbdr_taskq, 1, PI_NET, "%s: rbdr_taskq", 426 device_get_nameunit(nic->dev)); 427 428 return (0); 429 } 430 431 /* Free RBDR ring and its receive buffers */ 432 static void 433 nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr) 434 { 435 struct mbuf *mbuf; 436 struct queue_set *qs; 437 struct rbdr_entry_t *desc; 438 struct rbuf_info *rinfo; 439 bus_addr_t buf_addr; 440 int head, tail, idx; 441 int err; 442 443 qs = nic->qs; 444 445 if ((qs == NULL) || (rbdr == NULL)) 446 return; 447 448 rbdr->enable = FALSE; 449 if (rbdr->rbdr_taskq != NULL) { 450 /* Remove tasks */ 451 while (taskqueue_cancel(rbdr->rbdr_taskq, 452 &rbdr->rbdr_task_nowait, NULL) != 0) { 453 /* Finish the nowait task first */ 454 taskqueue_drain(rbdr->rbdr_taskq, 455 &rbdr->rbdr_task_nowait); 456 } 457 taskqueue_free(rbdr->rbdr_taskq); 458 rbdr->rbdr_taskq = NULL; 459 460 while (taskqueue_cancel(taskqueue_thread, 461 &rbdr->rbdr_task, NULL) != 0) { 462 /* Now finish the sleepable task */ 463 taskqueue_drain(taskqueue_thread, &rbdr->rbdr_task); 464 } 465 } 466 467 /* 468 * Free all of the memory under the RB descriptors. 469 * There are assumptions here: 470 * 1. Corresponding RBDR is disabled 471 * - it is safe to operate using head and tail indexes 472 * 2. All bffers that were received are properly freed by 473 * the receive handler 474 * - there is no need to unload DMA map and free MBUF for other 475 * descriptors than unused ones 476 */ 477 if (rbdr->rbdr_buff_dmat != NULL) { 478 head = rbdr->head; 479 tail = rbdr->tail; 480 while (head != tail) { 481 desc = GET_RBDR_DESC(rbdr, head); 482 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN; 483 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr)); 484 bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap); 485 mbuf = rinfo->mbuf; 486 /* This will destroy everything including rinfo! */ 487 m_freem(mbuf); 488 head++; 489 head &= (rbdr->dmem.q_len - 1); 490 } 491 /* Free tail descriptor */ 492 desc = GET_RBDR_DESC(rbdr, tail); 493 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN; 494 rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr)); 495 bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap); 496 mbuf = rinfo->mbuf; 497 /* This will destroy everything including rinfo! */ 498 m_freem(mbuf); 499 500 /* Destroy DMA maps */ 501 for (idx = 0; idx < qs->rbdr_len; idx++) { 502 if (rbdr->rbdr_buff_dmaps[idx] == NULL) 503 continue; 504 err = bus_dmamap_destroy(rbdr->rbdr_buff_dmat, 505 rbdr->rbdr_buff_dmaps[idx]); 506 KASSERT(err == 0, 507 ("%s: Could not destroy DMA map for RB, desc: %d", 508 __func__, idx)); 509 rbdr->rbdr_buff_dmaps[idx] = NULL; 510 } 511 512 /* Now destroy the tag */ 513 err = bus_dma_tag_destroy(rbdr->rbdr_buff_dmat); 514 KASSERT(err == 0, 515 ("%s: Trying to destroy BUSY DMA tag", __func__)); 516 517 rbdr->head = 0; 518 rbdr->tail = 0; 519 } 520 521 /* Free RBDR ring */ 522 nicvf_free_q_desc_mem(nic, &rbdr->dmem); 523 } 524 525 /* 526 * Refill receive buffer descriptors with new buffers. 527 */ 528 static int 529 nicvf_refill_rbdr(struct rbdr *rbdr, int mflags) 530 { 531 struct nicvf *nic; 532 struct queue_set *qs; 533 int rbdr_idx; 534 int tail, qcount; 535 int refill_rb_cnt; 536 struct rbdr_entry_t *desc; 537 bus_dmamap_t dmap; 538 bus_addr_t rbuf; 539 boolean_t rb_alloc_fail; 540 int new_rb; 541 542 rb_alloc_fail = TRUE; 543 new_rb = 0; 544 nic = rbdr->nic; 545 qs = nic->qs; 546 rbdr_idx = rbdr->idx; 547 548 /* Check if it's enabled */ 549 if (!rbdr->enable) 550 return (0); 551 552 /* Get no of desc's to be refilled */ 553 qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx); 554 qcount &= 0x7FFFF; 555 /* Doorbell can be ringed with a max of ring size minus 1 */ 556 if (qcount >= (qs->rbdr_len - 1)) { 557 rb_alloc_fail = FALSE; 558 goto out; 559 } else 560 refill_rb_cnt = qs->rbdr_len - qcount - 1; 561 562 /* Start filling descs from tail */ 563 tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3; 564 while (refill_rb_cnt) { 565 tail++; 566 tail &= (rbdr->dmem.q_len - 1); 567 568 dmap = rbdr->rbdr_buff_dmaps[tail]; 569 if (nicvf_alloc_rcv_buffer(nic, rbdr, dmap, mflags, 570 DMA_BUFFER_LEN, &rbuf)) { 571 /* Something went wrong. Resign */ 572 break; 573 } 574 desc = GET_RBDR_DESC(rbdr, tail); 575 desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN); 576 refill_rb_cnt--; 577 new_rb++; 578 } 579 580 /* make sure all memory stores are done before ringing doorbell */ 581 wmb(); 582 583 /* Check if buffer allocation failed */ 584 if (refill_rb_cnt == 0) 585 rb_alloc_fail = FALSE; 586 587 /* Notify HW */ 588 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, 589 rbdr_idx, new_rb); 590 out: 591 if (!rb_alloc_fail) { 592 /* 593 * Re-enable RBDR interrupts only 594 * if buffer allocation is success. 595 */ 596 nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx); 597 598 return (0); 599 } 600 601 return (ENOMEM); 602 } 603 604 /* Refill RBs even if sleep is needed to reclaim memory */ 605 static void 606 nicvf_rbdr_task(void *arg, int pending) 607 { 608 struct rbdr *rbdr; 609 int err; 610 611 rbdr = (struct rbdr *)arg; 612 613 err = nicvf_refill_rbdr(rbdr, M_WAITOK); 614 if (__predict_false(err != 0)) { 615 panic("%s: Failed to refill RBs even when sleep enabled", 616 __func__); 617 } 618 } 619 620 /* Refill RBs as soon as possible without waiting */ 621 static void 622 nicvf_rbdr_task_nowait(void *arg, int pending) 623 { 624 struct rbdr *rbdr; 625 int err; 626 627 rbdr = (struct rbdr *)arg; 628 629 err = nicvf_refill_rbdr(rbdr, M_NOWAIT); 630 if (err != 0) { 631 /* 632 * Schedule another, sleepable kernel thread 633 * that will for sure refill the buffers. 634 */ 635 taskqueue_enqueue(taskqueue_thread, &rbdr->rbdr_task); 636 } 637 } 638 639 static int 640 nicvf_rcv_pkt_handler(struct nicvf *nic, struct cmp_queue *cq, 641 struct cqe_rx_t *cqe_rx, int cqe_type) 642 { 643 struct mbuf *mbuf; 644 struct rcv_queue *rq; 645 int rq_idx; 646 int err = 0; 647 648 rq_idx = cqe_rx->rq_idx; 649 rq = &nic->qs->rq[rq_idx]; 650 651 /* Check for errors */ 652 err = nicvf_check_cqe_rx_errs(nic, cq, cqe_rx); 653 if (err && !cqe_rx->rb_cnt) 654 return (0); 655 656 mbuf = nicvf_get_rcv_mbuf(nic, cqe_rx); 657 if (mbuf == NULL) { 658 dprintf(nic->dev, "Packet not received\n"); 659 return (0); 660 } 661 662 /* If error packet */ 663 if (err != 0) { 664 m_freem(mbuf); 665 return (0); 666 } 667 668 if (rq->lro_enabled && 669 ((cqe_rx->l3_type == L3TYPE_IPV4) && (cqe_rx->l4_type == L4TYPE_TCP)) && 670 (mbuf->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) == 671 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) { 672 /* 673 * At this point it is known that there are no errors in the 674 * packet. Attempt to LRO enqueue. Send to stack if no resources 675 * or enqueue error. 676 */ 677 if ((rq->lro.lro_cnt != 0) && 678 (tcp_lro_rx(&rq->lro, mbuf, 0) == 0)) 679 return (0); 680 } 681 /* 682 * Push this packet to the stack later to avoid 683 * unlocking completion task in the middle of work. 684 */ 685 err = buf_ring_enqueue(cq->rx_br, mbuf); 686 if (err != 0) { 687 /* 688 * Failed to enqueue this mbuf. 689 * We don't drop it, just schedule another task. 690 */ 691 return (err); 692 } 693 694 return (0); 695 } 696 697 static void 698 nicvf_snd_pkt_handler(struct nicvf *nic, struct cmp_queue *cq, 699 struct cqe_send_t *cqe_tx, int cqe_type) 700 { 701 bus_dmamap_t dmap; 702 struct mbuf *mbuf; 703 struct snd_queue *sq; 704 struct sq_hdr_subdesc *hdr; 705 706 mbuf = NULL; 707 sq = &nic->qs->sq[cqe_tx->sq_idx]; 708 709 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, cqe_tx->sqe_ptr); 710 if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) 711 return; 712 713 dprintf(nic->dev, 714 "%s Qset #%d SQ #%d SQ ptr #%d subdesc count %d\n", 715 __func__, cqe_tx->sq_qs, cqe_tx->sq_idx, 716 cqe_tx->sqe_ptr, hdr->subdesc_cnt); 717 718 dmap = (bus_dmamap_t)sq->snd_buff[cqe_tx->sqe_ptr].dmap; 719 bus_dmamap_unload(sq->snd_buff_dmat, dmap); 720 721 mbuf = (struct mbuf *)sq->snd_buff[cqe_tx->sqe_ptr].mbuf; 722 if (mbuf != NULL) { 723 m_freem(mbuf); 724 sq->snd_buff[cqe_tx->sqe_ptr].mbuf = NULL; 725 nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1); 726 } 727 728 nicvf_check_cqe_tx_errs(nic, cq, cqe_tx); 729 } 730 731 static int 732 nicvf_cq_intr_handler(struct nicvf *nic, uint8_t cq_idx) 733 { 734 struct mbuf *mbuf; 735 struct ifnet *ifp; 736 int processed_cqe, work_done = 0, tx_done = 0; 737 int cqe_count, cqe_head; 738 struct queue_set *qs = nic->qs; 739 struct cmp_queue *cq = &qs->cq[cq_idx]; 740 struct snd_queue *sq = &qs->sq[cq_idx]; 741 struct rcv_queue *rq; 742 struct cqe_rx_t *cq_desc; 743 struct lro_ctrl *lro; 744 int rq_idx; 745 int cmp_err; 746 747 NICVF_CMP_LOCK(cq); 748 cmp_err = 0; 749 processed_cqe = 0; 750 /* Get no of valid CQ entries to process */ 751 cqe_count = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS, cq_idx); 752 cqe_count &= CQ_CQE_COUNT; 753 if (cqe_count == 0) 754 goto out; 755 756 /* Get head of the valid CQ entries */ 757 cqe_head = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_HEAD, cq_idx) >> 9; 758 cqe_head &= 0xFFFF; 759 760 dprintf(nic->dev, "%s CQ%d cqe_count %d cqe_head %d\n", 761 __func__, cq_idx, cqe_count, cqe_head); 762 while (processed_cqe < cqe_count) { 763 /* Get the CQ descriptor */ 764 cq_desc = (struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head); 765 cqe_head++; 766 cqe_head &= (cq->dmem.q_len - 1); 767 /* Prefetch next CQ descriptor */ 768 __builtin_prefetch((struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head)); 769 770 dprintf(nic->dev, "CQ%d cq_desc->cqe_type %d\n", cq_idx, 771 cq_desc->cqe_type); 772 switch (cq_desc->cqe_type) { 773 case CQE_TYPE_RX: 774 cmp_err = nicvf_rcv_pkt_handler(nic, cq, cq_desc, 775 CQE_TYPE_RX); 776 if (__predict_false(cmp_err != 0)) { 777 /* 778 * Ups. Cannot finish now. 779 * Let's try again later. 780 */ 781 goto done; 782 } 783 work_done++; 784 break; 785 case CQE_TYPE_SEND: 786 nicvf_snd_pkt_handler(nic, cq, (void *)cq_desc, 787 CQE_TYPE_SEND); 788 tx_done++; 789 break; 790 case CQE_TYPE_INVALID: 791 case CQE_TYPE_RX_SPLIT: 792 case CQE_TYPE_RX_TCP: 793 case CQE_TYPE_SEND_PTP: 794 /* Ignore for now */ 795 break; 796 } 797 processed_cqe++; 798 } 799 done: 800 dprintf(nic->dev, 801 "%s CQ%d processed_cqe %d work_done %d\n", 802 __func__, cq_idx, processed_cqe, work_done); 803 804 /* Ring doorbell to inform H/W to reuse processed CQEs */ 805 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_DOOR, cq_idx, processed_cqe); 806 807 if ((tx_done > 0) && 808 ((if_getdrvflags(nic->ifp) & IFF_DRV_RUNNING) != 0)) { 809 /* Reenable TXQ if its stopped earlier due to SQ full */ 810 if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 811 taskqueue_enqueue(sq->snd_taskq, &sq->snd_task); 812 } 813 out: 814 /* 815 * Flush any outstanding LRO work 816 */ 817 rq_idx = cq_idx; 818 rq = &nic->qs->rq[rq_idx]; 819 lro = &rq->lro; 820 tcp_lro_flush_all(lro); 821 822 NICVF_CMP_UNLOCK(cq); 823 824 ifp = nic->ifp; 825 /* Push received MBUFs to the stack */ 826 while (!buf_ring_empty(cq->rx_br)) { 827 mbuf = buf_ring_dequeue_mc(cq->rx_br); 828 if (__predict_true(mbuf != NULL)) 829 (*ifp->if_input)(ifp, mbuf); 830 } 831 832 return (cmp_err); 833 } 834 835 /* 836 * Qset error interrupt handler 837 * 838 * As of now only CQ errors are handled 839 */ 840 static void 841 nicvf_qs_err_task(void *arg, int pending) 842 { 843 struct nicvf *nic; 844 struct queue_set *qs; 845 int qidx; 846 uint64_t status; 847 boolean_t enable = TRUE; 848 849 nic = (struct nicvf *)arg; 850 qs = nic->qs; 851 852 /* Deactivate network interface */ 853 if_setdrvflagbits(nic->ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 854 855 /* Check if it is CQ err */ 856 for (qidx = 0; qidx < qs->cq_cnt; qidx++) { 857 status = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS, 858 qidx); 859 if ((status & CQ_ERR_MASK) == 0) 860 continue; 861 /* Process already queued CQEs and reconfig CQ */ 862 nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx); 863 nicvf_sq_disable(nic, qidx); 864 (void)nicvf_cq_intr_handler(nic, qidx); 865 nicvf_cmp_queue_config(nic, qs, qidx, enable); 866 nicvf_sq_free_used_descs(nic, &qs->sq[qidx], qidx); 867 nicvf_sq_enable(nic, &qs->sq[qidx], qidx); 868 nicvf_enable_intr(nic, NICVF_INTR_CQ, qidx); 869 } 870 871 if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 872 /* Re-enable Qset error interrupt */ 873 nicvf_enable_intr(nic, NICVF_INTR_QS_ERR, 0); 874 } 875 876 static void 877 nicvf_cmp_task(void *arg, int pending) 878 { 879 struct cmp_queue *cq; 880 struct nicvf *nic; 881 int cmp_err; 882 883 cq = (struct cmp_queue *)arg; 884 nic = cq->nic; 885 886 /* Handle CQ descriptors */ 887 cmp_err = nicvf_cq_intr_handler(nic, cq->idx); 888 if (__predict_false(cmp_err != 0)) { 889 /* 890 * Schedule another thread here since we did not 891 * process the entire CQ due to Tx or Rx CQ parse error. 892 */ 893 taskqueue_enqueue(cq->cmp_taskq, &cq->cmp_task); 894 } 895 896 nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx); 897 /* Reenable interrupt (previously disabled in nicvf_intr_handler() */ 898 nicvf_enable_intr(nic, NICVF_INTR_CQ, cq->idx); 899 900 } 901 902 /* Initialize completion queue */ 903 static int 904 nicvf_init_cmp_queue(struct nicvf *nic, struct cmp_queue *cq, int q_len, 905 int qidx) 906 { 907 int err; 908 909 /* Initizalize lock */ 910 snprintf(cq->mtx_name, sizeof(cq->mtx_name), "%s: CQ(%d) lock", 911 device_get_nameunit(nic->dev), qidx); 912 mtx_init(&cq->mtx, cq->mtx_name, NULL, MTX_DEF); 913 914 err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE, 915 NICVF_CQ_BASE_ALIGN_BYTES); 916 917 if (err != 0) { 918 device_printf(nic->dev, 919 "Could not allocate DMA memory for CQ\n"); 920 return (err); 921 } 922 923 cq->desc = cq->dmem.base; 924 cq->thresh = pass1_silicon(nic->dev) ? 0 : CMP_QUEUE_CQE_THRESH; 925 cq->nic = nic; 926 cq->idx = qidx; 927 nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1; 928 929 cq->rx_br = buf_ring_alloc(CMP_QUEUE_LEN * 8, M_DEVBUF, M_WAITOK, 930 &cq->mtx); 931 932 /* Allocate taskqueue */ 933 NET_TASK_INIT(&cq->cmp_task, 0, nicvf_cmp_task, cq); 934 cq->cmp_taskq = taskqueue_create_fast("nicvf_cmp_taskq", M_WAITOK, 935 taskqueue_thread_enqueue, &cq->cmp_taskq); 936 taskqueue_start_threads(&cq->cmp_taskq, 1, PI_NET, "%s: cmp_taskq(%d)", 937 device_get_nameunit(nic->dev), qidx); 938 939 return (0); 940 } 941 942 static void 943 nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq) 944 { 945 946 if (cq == NULL) 947 return; 948 /* 949 * The completion queue itself should be disabled by now 950 * (ref. nicvf_snd_queue_config()). 951 * Ensure that it is safe to disable it or panic. 952 */ 953 if (cq->enable) 954 panic("%s: Trying to free working CQ(%d)", __func__, cq->idx); 955 956 if (cq->cmp_taskq != NULL) { 957 /* Remove task */ 958 while (taskqueue_cancel(cq->cmp_taskq, &cq->cmp_task, NULL) != 0) 959 taskqueue_drain(cq->cmp_taskq, &cq->cmp_task); 960 961 taskqueue_free(cq->cmp_taskq); 962 cq->cmp_taskq = NULL; 963 } 964 /* 965 * Completion interrupt will possibly enable interrupts again 966 * so disable interrupting now after we finished processing 967 * completion task. It is safe to do so since the corresponding CQ 968 * was already disabled. 969 */ 970 nicvf_disable_intr(nic, NICVF_INTR_CQ, cq->idx); 971 nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx); 972 973 NICVF_CMP_LOCK(cq); 974 nicvf_free_q_desc_mem(nic, &cq->dmem); 975 drbr_free(cq->rx_br, M_DEVBUF); 976 NICVF_CMP_UNLOCK(cq); 977 mtx_destroy(&cq->mtx); 978 memset(cq->mtx_name, 0, sizeof(cq->mtx_name)); 979 } 980 981 int 982 nicvf_xmit_locked(struct snd_queue *sq) 983 { 984 struct nicvf *nic; 985 struct ifnet *ifp; 986 struct mbuf *next; 987 int err; 988 989 NICVF_TX_LOCK_ASSERT(sq); 990 991 nic = sq->nic; 992 ifp = nic->ifp; 993 err = 0; 994 995 while ((next = drbr_peek(ifp, sq->br)) != NULL) { 996 /* Send a copy of the frame to the BPF listener */ 997 ETHER_BPF_MTAP(ifp, next); 998 999 err = nicvf_tx_mbuf_locked(sq, &next); 1000 if (err != 0) { 1001 if (next == NULL) 1002 drbr_advance(ifp, sq->br); 1003 else 1004 drbr_putback(ifp, sq->br, next); 1005 1006 break; 1007 } 1008 drbr_advance(ifp, sq->br); 1009 } 1010 return (err); 1011 } 1012 1013 static void 1014 nicvf_snd_task(void *arg, int pending) 1015 { 1016 struct snd_queue *sq = (struct snd_queue *)arg; 1017 struct nicvf *nic; 1018 struct ifnet *ifp; 1019 int err; 1020 1021 nic = sq->nic; 1022 ifp = nic->ifp; 1023 1024 /* 1025 * Skip sending anything if the driver is not running, 1026 * SQ full or link is down. 1027 */ 1028 if (((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != 1029 IFF_DRV_RUNNING) || !nic->link_up) 1030 return; 1031 1032 NICVF_TX_LOCK(sq); 1033 err = nicvf_xmit_locked(sq); 1034 NICVF_TX_UNLOCK(sq); 1035 /* Try again */ 1036 if (err != 0) 1037 taskqueue_enqueue(sq->snd_taskq, &sq->snd_task); 1038 } 1039 1040 /* Initialize transmit queue */ 1041 static int 1042 nicvf_init_snd_queue(struct nicvf *nic, struct snd_queue *sq, int q_len, 1043 int qidx) 1044 { 1045 size_t i; 1046 int err; 1047 1048 /* Initizalize TX lock for this queue */ 1049 snprintf(sq->mtx_name, sizeof(sq->mtx_name), "%s: SQ(%d) lock", 1050 device_get_nameunit(nic->dev), qidx); 1051 mtx_init(&sq->mtx, sq->mtx_name, NULL, MTX_DEF); 1052 1053 NICVF_TX_LOCK(sq); 1054 /* Allocate buffer ring */ 1055 sq->br = buf_ring_alloc(q_len / MIN_SQ_DESC_PER_PKT_XMIT, M_DEVBUF, 1056 M_NOWAIT, &sq->mtx); 1057 if (sq->br == NULL) { 1058 device_printf(nic->dev, 1059 "ERROR: Could not set up buf ring for SQ(%d)\n", qidx); 1060 err = ENOMEM; 1061 goto error; 1062 } 1063 1064 /* Allocate DMA memory for Tx descriptors */ 1065 err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE, 1066 NICVF_SQ_BASE_ALIGN_BYTES); 1067 if (err != 0) { 1068 device_printf(nic->dev, 1069 "Could not allocate DMA memory for SQ\n"); 1070 goto error; 1071 } 1072 1073 sq->desc = sq->dmem.base; 1074 sq->head = sq->tail = 0; 1075 atomic_store_rel_int(&sq->free_cnt, q_len - 1); 1076 sq->thresh = SND_QUEUE_THRESH; 1077 sq->idx = qidx; 1078 sq->nic = nic; 1079 1080 /* 1081 * Allocate DMA maps for Tx buffers 1082 */ 1083 1084 /* Create DMA tag first */ 1085 err = bus_dma_tag_create( 1086 bus_get_dma_tag(nic->dev), /* parent tag */ 1087 1, /* alignment */ 1088 0, /* boundary */ 1089 BUS_SPACE_MAXADDR, /* lowaddr */ 1090 BUS_SPACE_MAXADDR, /* highaddr */ 1091 NULL, NULL, /* filtfunc, filtfuncarg */ 1092 NICVF_TSO_MAXSIZE, /* maxsize */ 1093 NICVF_TSO_NSEGS, /* nsegments */ 1094 MCLBYTES, /* maxsegsize */ 1095 0, /* flags */ 1096 NULL, NULL, /* lockfunc, lockfuncarg */ 1097 &sq->snd_buff_dmat); /* dmat */ 1098 1099 if (err != 0) { 1100 device_printf(nic->dev, 1101 "Failed to create busdma tag for Tx buffers\n"); 1102 goto error; 1103 } 1104 1105 /* Allocate send buffers array */ 1106 sq->snd_buff = malloc(sizeof(*sq->snd_buff) * q_len, M_NICVF, 1107 (M_NOWAIT | M_ZERO)); 1108 if (sq->snd_buff == NULL) { 1109 device_printf(nic->dev, 1110 "Could not allocate memory for Tx buffers array\n"); 1111 err = ENOMEM; 1112 goto error; 1113 } 1114 1115 /* Now populate maps */ 1116 for (i = 0; i < q_len; i++) { 1117 err = bus_dmamap_create(sq->snd_buff_dmat, 0, 1118 &sq->snd_buff[i].dmap); 1119 if (err != 0) { 1120 device_printf(nic->dev, 1121 "Failed to create DMA maps for Tx buffers\n"); 1122 goto error; 1123 } 1124 } 1125 NICVF_TX_UNLOCK(sq); 1126 1127 /* Allocate taskqueue */ 1128 TASK_INIT(&sq->snd_task, 0, nicvf_snd_task, sq); 1129 sq->snd_taskq = taskqueue_create_fast("nicvf_snd_taskq", M_WAITOK, 1130 taskqueue_thread_enqueue, &sq->snd_taskq); 1131 taskqueue_start_threads(&sq->snd_taskq, 1, PI_NET, "%s: snd_taskq(%d)", 1132 device_get_nameunit(nic->dev), qidx); 1133 1134 return (0); 1135 error: 1136 NICVF_TX_UNLOCK(sq); 1137 return (err); 1138 } 1139 1140 static void 1141 nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq) 1142 { 1143 struct queue_set *qs = nic->qs; 1144 size_t i; 1145 int err; 1146 1147 if (sq == NULL) 1148 return; 1149 1150 if (sq->snd_taskq != NULL) { 1151 /* Remove task */ 1152 while (taskqueue_cancel(sq->snd_taskq, &sq->snd_task, NULL) != 0) 1153 taskqueue_drain(sq->snd_taskq, &sq->snd_task); 1154 1155 taskqueue_free(sq->snd_taskq); 1156 sq->snd_taskq = NULL; 1157 } 1158 1159 NICVF_TX_LOCK(sq); 1160 if (sq->snd_buff_dmat != NULL) { 1161 if (sq->snd_buff != NULL) { 1162 for (i = 0; i < qs->sq_len; i++) { 1163 m_freem(sq->snd_buff[i].mbuf); 1164 sq->snd_buff[i].mbuf = NULL; 1165 1166 bus_dmamap_unload(sq->snd_buff_dmat, 1167 sq->snd_buff[i].dmap); 1168 err = bus_dmamap_destroy(sq->snd_buff_dmat, 1169 sq->snd_buff[i].dmap); 1170 /* 1171 * If bus_dmamap_destroy fails it can cause 1172 * random panic later if the tag is also 1173 * destroyed in the process. 1174 */ 1175 KASSERT(err == 0, 1176 ("%s: Could not destroy DMA map for SQ", 1177 __func__)); 1178 } 1179 } 1180 1181 free(sq->snd_buff, M_NICVF); 1182 1183 err = bus_dma_tag_destroy(sq->snd_buff_dmat); 1184 KASSERT(err == 0, 1185 ("%s: Trying to destroy BUSY DMA tag", __func__)); 1186 } 1187 1188 /* Free private driver ring for this send queue */ 1189 if (sq->br != NULL) 1190 drbr_free(sq->br, M_DEVBUF); 1191 1192 if (sq->dmem.base != NULL) 1193 nicvf_free_q_desc_mem(nic, &sq->dmem); 1194 1195 NICVF_TX_UNLOCK(sq); 1196 /* Destroy Tx lock */ 1197 mtx_destroy(&sq->mtx); 1198 memset(sq->mtx_name, 0, sizeof(sq->mtx_name)); 1199 } 1200 1201 static void 1202 nicvf_reclaim_snd_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1203 { 1204 1205 /* Disable send queue */ 1206 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0); 1207 /* Check if SQ is stopped */ 1208 if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01)) 1209 return; 1210 /* Reset send queue */ 1211 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 1212 } 1213 1214 static void 1215 nicvf_reclaim_rcv_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1216 { 1217 union nic_mbx mbx = {}; 1218 1219 /* Make sure all packets in the pipeline are written back into mem */ 1220 mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC; 1221 nicvf_send_msg_to_pf(nic, &mbx); 1222 } 1223 1224 static void 1225 nicvf_reclaim_cmp_queue(struct nicvf *nic, struct queue_set *qs, int qidx) 1226 { 1227 1228 /* Disable timer threshold (doesn't get reset upon CQ reset */ 1229 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0); 1230 /* Disable completion queue */ 1231 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0); 1232 /* Reset completion queue */ 1233 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 1234 } 1235 1236 static void 1237 nicvf_reclaim_rbdr(struct nicvf *nic, struct rbdr *rbdr, int qidx) 1238 { 1239 uint64_t tmp, fifo_state; 1240 int timeout = 10; 1241 1242 /* Save head and tail pointers for feeing up buffers */ 1243 rbdr->head = 1244 nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_HEAD, qidx) >> 3; 1245 rbdr->tail = 1246 nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, qidx) >> 3; 1247 1248 /* 1249 * If RBDR FIFO is in 'FAIL' state then do a reset first 1250 * before relaiming. 1251 */ 1252 fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx); 1253 if (((fifo_state >> 62) & 0x03) == 0x3) { 1254 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, 1255 qidx, NICVF_RBDR_RESET); 1256 } 1257 1258 /* Disable RBDR */ 1259 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0); 1260 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 1261 return; 1262 while (1) { 1263 tmp = nicvf_queue_reg_read(nic, 1264 NIC_QSET_RBDR_0_1_PREFETCH_STATUS, qidx); 1265 if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF)) 1266 break; 1267 1268 DELAY(1000); 1269 timeout--; 1270 if (!timeout) { 1271 device_printf(nic->dev, 1272 "Failed polling on prefetch status\n"); 1273 return; 1274 } 1275 } 1276 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 1277 NICVF_RBDR_RESET); 1278 1279 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02)) 1280 return; 1281 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00); 1282 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 1283 return; 1284 } 1285 1286 /* Configures receive queue */ 1287 static void 1288 nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs, 1289 int qidx, bool enable) 1290 { 1291 union nic_mbx mbx = {}; 1292 struct rcv_queue *rq; 1293 struct rq_cfg rq_cfg; 1294 struct ifnet *ifp; 1295 struct lro_ctrl *lro; 1296 1297 ifp = nic->ifp; 1298 1299 rq = &qs->rq[qidx]; 1300 rq->enable = enable; 1301 1302 lro = &rq->lro; 1303 1304 /* Disable receive queue */ 1305 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0); 1306 1307 if (!rq->enable) { 1308 nicvf_reclaim_rcv_queue(nic, qs, qidx); 1309 /* Free LRO memory */ 1310 tcp_lro_free(lro); 1311 rq->lro_enabled = FALSE; 1312 return; 1313 } 1314 1315 /* Configure LRO if enabled */ 1316 rq->lro_enabled = FALSE; 1317 if ((if_getcapenable(ifp) & IFCAP_LRO) != 0) { 1318 if (tcp_lro_init(lro) != 0) { 1319 device_printf(nic->dev, 1320 "Failed to initialize LRO for RXQ%d\n", qidx); 1321 } else { 1322 rq->lro_enabled = TRUE; 1323 lro->ifp = nic->ifp; 1324 } 1325 } 1326 1327 rq->cq_qs = qs->vnic_id; 1328 rq->cq_idx = qidx; 1329 rq->start_rbdr_qs = qs->vnic_id; 1330 rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1; 1331 rq->cont_rbdr_qs = qs->vnic_id; 1332 rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1; 1333 /* all writes of RBDR data to be loaded into L2 Cache as well*/ 1334 rq->caching = 1; 1335 1336 /* Send a mailbox msg to PF to config RQ */ 1337 mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG; 1338 mbx.rq.qs_num = qs->vnic_id; 1339 mbx.rq.rq_num = qidx; 1340 mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) | 1341 (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) | 1342 (rq->cont_qs_rbdr_idx << 8) | (rq->start_rbdr_qs << 1) | 1343 (rq->start_qs_rbdr_idx); 1344 nicvf_send_msg_to_pf(nic, &mbx); 1345 1346 mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG; 1347 mbx.rq.cfg = (1UL << 63) | (1UL << 62) | (qs->vnic_id << 0); 1348 nicvf_send_msg_to_pf(nic, &mbx); 1349 1350 /* 1351 * RQ drop config 1352 * Enable CQ drop to reserve sufficient CQEs for all tx packets 1353 */ 1354 mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG; 1355 mbx.rq.cfg = (1UL << 62) | (RQ_CQ_DROP << 8); 1356 nicvf_send_msg_to_pf(nic, &mbx); 1357 1358 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, 0x00); 1359 1360 /* Enable Receive queue */ 1361 rq_cfg.ena = 1; 1362 rq_cfg.tcp_ena = 0; 1363 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 1364 *(uint64_t *)&rq_cfg); 1365 } 1366 1367 /* Configures completion queue */ 1368 static void 1369 nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs, 1370 int qidx, boolean_t enable) 1371 { 1372 struct cmp_queue *cq; 1373 struct cq_cfg cq_cfg; 1374 1375 cq = &qs->cq[qidx]; 1376 cq->enable = enable; 1377 1378 if (!cq->enable) { 1379 nicvf_reclaim_cmp_queue(nic, qs, qidx); 1380 return; 1381 } 1382 1383 /* Reset completion queue */ 1384 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 1385 1386 /* Set completion queue base address */ 1387 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE, qidx, 1388 (uint64_t)(cq->dmem.phys_base)); 1389 1390 /* Enable Completion queue */ 1391 cq_cfg.ena = 1; 1392 cq_cfg.reset = 0; 1393 cq_cfg.caching = 0; 1394 cq_cfg.qsize = CMP_QSIZE; 1395 cq_cfg.avg_con = 0; 1396 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(uint64_t *)&cq_cfg); 1397 1398 /* Set threshold value for interrupt generation */ 1399 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh); 1400 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 1401 nic->cq_coalesce_usecs); 1402 } 1403 1404 /* Configures transmit queue */ 1405 static void 1406 nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs, int qidx, 1407 boolean_t enable) 1408 { 1409 union nic_mbx mbx = {}; 1410 struct snd_queue *sq; 1411 struct sq_cfg sq_cfg; 1412 1413 sq = &qs->sq[qidx]; 1414 sq->enable = enable; 1415 1416 if (!sq->enable) { 1417 nicvf_reclaim_snd_queue(nic, qs, qidx); 1418 return; 1419 } 1420 1421 /* Reset send queue */ 1422 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 1423 1424 sq->cq_qs = qs->vnic_id; 1425 sq->cq_idx = qidx; 1426 1427 /* Send a mailbox msg to PF to config SQ */ 1428 mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG; 1429 mbx.sq.qs_num = qs->vnic_id; 1430 mbx.sq.sq_num = qidx; 1431 mbx.sq.sqs_mode = nic->sqs_mode; 1432 mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx; 1433 nicvf_send_msg_to_pf(nic, &mbx); 1434 1435 /* Set queue base address */ 1436 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE, qidx, 1437 (uint64_t)(sq->dmem.phys_base)); 1438 1439 /* Enable send queue & set queue size */ 1440 sq_cfg.ena = 1; 1441 sq_cfg.reset = 0; 1442 sq_cfg.ldwb = 0; 1443 sq_cfg.qsize = SND_QSIZE; 1444 sq_cfg.tstmp_bgx_intf = 0; 1445 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(uint64_t *)&sq_cfg); 1446 1447 /* Set threshold value for interrupt generation */ 1448 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh); 1449 } 1450 1451 /* Configures receive buffer descriptor ring */ 1452 static void 1453 nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs, int qidx, 1454 boolean_t enable) 1455 { 1456 struct rbdr *rbdr; 1457 struct rbdr_cfg rbdr_cfg; 1458 1459 rbdr = &qs->rbdr[qidx]; 1460 nicvf_reclaim_rbdr(nic, rbdr, qidx); 1461 if (!enable) 1462 return; 1463 1464 /* Set descriptor base address */ 1465 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE, qidx, 1466 (uint64_t)(rbdr->dmem.phys_base)); 1467 1468 /* Enable RBDR & set queue size */ 1469 /* Buffer size should be in multiples of 128 bytes */ 1470 rbdr_cfg.ena = 1; 1471 rbdr_cfg.reset = 0; 1472 rbdr_cfg.ldwb = 0; 1473 rbdr_cfg.qsize = RBDR_SIZE; 1474 rbdr_cfg.avg_con = 0; 1475 rbdr_cfg.lines = rbdr->dma_size / 128; 1476 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 1477 *(uint64_t *)&rbdr_cfg); 1478 1479 /* Notify HW */ 1480 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, qidx, 1481 qs->rbdr_len - 1); 1482 1483 /* Set threshold value for interrupt generation */ 1484 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH, qidx, 1485 rbdr->thresh - 1); 1486 } 1487 1488 /* Requests PF to assign and enable Qset */ 1489 void 1490 nicvf_qset_config(struct nicvf *nic, boolean_t enable) 1491 { 1492 union nic_mbx mbx = {}; 1493 struct queue_set *qs; 1494 struct qs_cfg *qs_cfg; 1495 1496 qs = nic->qs; 1497 if (qs == NULL) { 1498 device_printf(nic->dev, 1499 "Qset is still not allocated, don't init queues\n"); 1500 return; 1501 } 1502 1503 qs->enable = enable; 1504 qs->vnic_id = nic->vf_id; 1505 1506 /* Send a mailbox msg to PF to config Qset */ 1507 mbx.qs.msg = NIC_MBOX_MSG_QS_CFG; 1508 mbx.qs.num = qs->vnic_id; 1509 1510 mbx.qs.cfg = 0; 1511 qs_cfg = (struct qs_cfg *)&mbx.qs.cfg; 1512 if (qs->enable) { 1513 qs_cfg->ena = 1; 1514 qs_cfg->vnic = qs->vnic_id; 1515 } 1516 nicvf_send_msg_to_pf(nic, &mbx); 1517 } 1518 1519 static void 1520 nicvf_free_resources(struct nicvf *nic) 1521 { 1522 int qidx; 1523 struct queue_set *qs; 1524 1525 qs = nic->qs; 1526 /* 1527 * Remove QS error task first since it has to be dead 1528 * to safely free completion queue tasks. 1529 */ 1530 if (qs->qs_err_taskq != NULL) { 1531 /* Shut down QS error tasks */ 1532 while (taskqueue_cancel(qs->qs_err_taskq, 1533 &qs->qs_err_task, NULL) != 0) { 1534 taskqueue_drain(qs->qs_err_taskq, &qs->qs_err_task); 1535 } 1536 taskqueue_free(qs->qs_err_taskq); 1537 qs->qs_err_taskq = NULL; 1538 } 1539 /* Free receive buffer descriptor ring */ 1540 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1541 nicvf_free_rbdr(nic, &qs->rbdr[qidx]); 1542 1543 /* Free completion queue */ 1544 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1545 nicvf_free_cmp_queue(nic, &qs->cq[qidx]); 1546 1547 /* Free send queue */ 1548 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1549 nicvf_free_snd_queue(nic, &qs->sq[qidx]); 1550 } 1551 1552 static int 1553 nicvf_alloc_resources(struct nicvf *nic) 1554 { 1555 struct queue_set *qs = nic->qs; 1556 int qidx; 1557 1558 /* Alloc receive buffer descriptor ring */ 1559 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) { 1560 if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len, 1561 DMA_BUFFER_LEN, qidx)) 1562 goto alloc_fail; 1563 } 1564 1565 /* Alloc send queue */ 1566 for (qidx = 0; qidx < qs->sq_cnt; qidx++) { 1567 if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx)) 1568 goto alloc_fail; 1569 } 1570 1571 /* Alloc completion queue */ 1572 for (qidx = 0; qidx < qs->cq_cnt; qidx++) { 1573 if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len, qidx)) 1574 goto alloc_fail; 1575 } 1576 1577 /* Allocate QS error taskqueue */ 1578 NET_TASK_INIT(&qs->qs_err_task, 0, nicvf_qs_err_task, nic); 1579 qs->qs_err_taskq = taskqueue_create_fast("nicvf_qs_err_taskq", M_WAITOK, 1580 taskqueue_thread_enqueue, &qs->qs_err_taskq); 1581 taskqueue_start_threads(&qs->qs_err_taskq, 1, PI_NET, "%s: qs_taskq", 1582 device_get_nameunit(nic->dev)); 1583 1584 return (0); 1585 alloc_fail: 1586 nicvf_free_resources(nic); 1587 return (ENOMEM); 1588 } 1589 1590 int 1591 nicvf_set_qset_resources(struct nicvf *nic) 1592 { 1593 struct queue_set *qs; 1594 1595 qs = malloc(sizeof(*qs), M_NICVF, (M_ZERO | M_WAITOK)); 1596 nic->qs = qs; 1597 1598 /* Set count of each queue */ 1599 qs->rbdr_cnt = RBDR_CNT; 1600 qs->rq_cnt = RCV_QUEUE_CNT; 1601 1602 qs->sq_cnt = SND_QUEUE_CNT; 1603 qs->cq_cnt = CMP_QUEUE_CNT; 1604 1605 /* Set queue lengths */ 1606 qs->rbdr_len = RCV_BUF_COUNT; 1607 qs->sq_len = SND_QUEUE_LEN; 1608 qs->cq_len = CMP_QUEUE_LEN; 1609 1610 nic->rx_queues = qs->rq_cnt; 1611 nic->tx_queues = qs->sq_cnt; 1612 1613 return (0); 1614 } 1615 1616 int 1617 nicvf_config_data_transfer(struct nicvf *nic, boolean_t enable) 1618 { 1619 boolean_t disable = FALSE; 1620 struct queue_set *qs; 1621 int qidx; 1622 1623 qs = nic->qs; 1624 if (qs == NULL) 1625 return (0); 1626 1627 if (enable) { 1628 if (nicvf_alloc_resources(nic) != 0) 1629 return (ENOMEM); 1630 1631 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1632 nicvf_snd_queue_config(nic, qs, qidx, enable); 1633 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1634 nicvf_cmp_queue_config(nic, qs, qidx, enable); 1635 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1636 nicvf_rbdr_config(nic, qs, qidx, enable); 1637 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1638 nicvf_rcv_queue_config(nic, qs, qidx, enable); 1639 } else { 1640 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1641 nicvf_rcv_queue_config(nic, qs, qidx, disable); 1642 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1643 nicvf_rbdr_config(nic, qs, qidx, disable); 1644 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1645 nicvf_snd_queue_config(nic, qs, qidx, disable); 1646 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1647 nicvf_cmp_queue_config(nic, qs, qidx, disable); 1648 1649 nicvf_free_resources(nic); 1650 } 1651 1652 return (0); 1653 } 1654 1655 /* 1656 * Get a free desc from SQ 1657 * returns descriptor ponter & descriptor number 1658 */ 1659 static __inline int 1660 nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt) 1661 { 1662 int qentry; 1663 1664 qentry = sq->tail; 1665 atomic_subtract_int(&sq->free_cnt, desc_cnt); 1666 sq->tail += desc_cnt; 1667 sq->tail &= (sq->dmem.q_len - 1); 1668 1669 return (qentry); 1670 } 1671 1672 /* Free descriptor back to SQ for future use */ 1673 static void 1674 nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt) 1675 { 1676 1677 atomic_add_int(&sq->free_cnt, desc_cnt); 1678 sq->head += desc_cnt; 1679 sq->head &= (sq->dmem.q_len - 1); 1680 } 1681 1682 static __inline int 1683 nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry) 1684 { 1685 qentry++; 1686 qentry &= (sq->dmem.q_len - 1); 1687 return (qentry); 1688 } 1689 1690 static void 1691 nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx) 1692 { 1693 uint64_t sq_cfg; 1694 1695 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1696 sq_cfg |= NICVF_SQ_EN; 1697 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1698 /* Ring doorbell so that H/W restarts processing SQEs */ 1699 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0); 1700 } 1701 1702 static void 1703 nicvf_sq_disable(struct nicvf *nic, int qidx) 1704 { 1705 uint64_t sq_cfg; 1706 1707 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1708 sq_cfg &= ~NICVF_SQ_EN; 1709 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1710 } 1711 1712 static void 1713 nicvf_sq_free_used_descs(struct nicvf *nic, struct snd_queue *sq, int qidx) 1714 { 1715 uint64_t head; 1716 struct snd_buff *snd_buff; 1717 struct sq_hdr_subdesc *hdr; 1718 1719 NICVF_TX_LOCK(sq); 1720 head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, 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 */ 1742 static __inline int 1743 nicvf_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, proto; 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 poff = proto = -1; 1779 switch (etype) { 1780 #ifdef INET6 1781 case ETHERTYPE_IPV6: 1782 if (mbuf->m_len < ehdrlen + sizeof(struct ip6_hdr)) { 1783 mbuf = m_pullup(mbuf, ehdrlen +sizeof(struct ip6_hdr)); 1784 sq->snd_buff[qentry].mbuf = NULL; 1785 if (mbuf == NULL) 1786 return (ENOBUFS); 1787 } 1788 poff = ip6_lasthdr(mbuf, ehdrlen, IPPROTO_IPV6, &proto); 1789 if (poff < 0) 1790 return (ENOBUFS); 1791 poff += ehdrlen; 1792 break; 1793 #endif 1794 #ifdef INET 1795 case ETHERTYPE_IP: 1796 if (mbuf->m_len < ehdrlen + sizeof(struct ip)) { 1797 mbuf = m_pullup(mbuf, ehdrlen + sizeof(struct ip)); 1798 sq->snd_buff[qentry].mbuf = mbuf; 1799 if (mbuf == NULL) 1800 return (ENOBUFS); 1801 } 1802 if (mbuf->m_pkthdr.csum_flags & CSUM_IP) 1803 hdr->csum_l3 = 1; /* Enable IP csum calculation */ 1804 1805 ip = (struct ip *)(mbuf->m_data + ehdrlen); 1806 iphlen = ip->ip_hl << 2; 1807 poff = ehdrlen + iphlen; 1808 proto = ip->ip_p; 1809 break; 1810 #endif 1811 } 1812 1813 #if defined(INET6) || defined(INET) 1814 if (poff > 0 && mbuf->m_pkthdr.csum_flags != 0) { 1815 switch (proto) { 1816 case IPPROTO_TCP: 1817 if ((mbuf->m_pkthdr.csum_flags & CSUM_TCP) == 0) 1818 break; 1819 1820 if (mbuf->m_len < (poff + sizeof(struct tcphdr))) { 1821 mbuf = m_pullup(mbuf, poff + sizeof(struct tcphdr)); 1822 sq->snd_buff[qentry].mbuf = mbuf; 1823 if (mbuf == NULL) 1824 return (ENOBUFS); 1825 } 1826 hdr->csum_l4 = SEND_L4_CSUM_TCP; 1827 break; 1828 case IPPROTO_UDP: 1829 if ((mbuf->m_pkthdr.csum_flags & CSUM_UDP) == 0) 1830 break; 1831 1832 if (mbuf->m_len < (poff + sizeof(struct udphdr))) { 1833 mbuf = m_pullup(mbuf, poff + sizeof(struct udphdr)); 1834 sq->snd_buff[qentry].mbuf = mbuf; 1835 if (mbuf == NULL) 1836 return (ENOBUFS); 1837 } 1838 hdr->csum_l4 = SEND_L4_CSUM_UDP; 1839 break; 1840 case IPPROTO_SCTP: 1841 if ((mbuf->m_pkthdr.csum_flags & CSUM_SCTP) == 0) 1842 break; 1843 1844 if (mbuf->m_len < (poff + sizeof(struct sctphdr))) { 1845 mbuf = m_pullup(mbuf, poff + sizeof(struct sctphdr)); 1846 sq->snd_buff[qentry].mbuf = mbuf; 1847 if (mbuf == NULL) 1848 return (ENOBUFS); 1849 } 1850 hdr->csum_l4 = SEND_L4_CSUM_SCTP; 1851 break; 1852 default: 1853 break; 1854 } 1855 hdr->l3_offset = ehdrlen; 1856 hdr->l4_offset = poff; 1857 } 1858 1859 if ((mbuf->m_pkthdr.tso_segsz != 0) && nic->hw_tso) { 1860 th = (struct tcphdr *)((caddr_t)(mbuf->m_data + poff)); 1861 1862 hdr->tso = 1; 1863 hdr->tso_start = poff + (th->th_off * 4); 1864 hdr->tso_max_paysize = mbuf->m_pkthdr.tso_segsz; 1865 hdr->inner_l3_offset = ehdrlen - 2; 1866 nic->drv_stats.tx_tso++; 1867 } 1868 #endif 1869 1870 return (0); 1871 } 1872 1873 /* 1874 * SQ GATHER subdescriptor 1875 * Must follow HDR descriptor 1876 */ 1877 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry, 1878 int size, uint64_t data) 1879 { 1880 struct sq_gather_subdesc *gather; 1881 1882 qentry &= (sq->dmem.q_len - 1); 1883 gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry); 1884 1885 memset(gather, 0, SND_QUEUE_DESC_SIZE); 1886 gather->subdesc_type = SQ_DESC_TYPE_GATHER; 1887 gather->ld_type = NIC_SEND_LD_TYPE_E_LDD; 1888 gather->size = size; 1889 gather->addr = data; 1890 } 1891 1892 /* Put an mbuf to a SQ for packet transfer. */ 1893 static int 1894 nicvf_tx_mbuf_locked(struct snd_queue *sq, struct mbuf **mbufp) 1895 { 1896 bus_dma_segment_t segs[256]; 1897 struct snd_buff *snd_buff; 1898 size_t seg; 1899 int nsegs, qentry; 1900 int subdesc_cnt; 1901 int err; 1902 1903 NICVF_TX_LOCK_ASSERT(sq); 1904 1905 if (sq->free_cnt == 0) 1906 return (ENOBUFS); 1907 1908 snd_buff = &sq->snd_buff[sq->tail]; 1909 1910 err = bus_dmamap_load_mbuf_sg(sq->snd_buff_dmat, snd_buff->dmap, 1911 *mbufp, segs, &nsegs, BUS_DMA_NOWAIT); 1912 if (__predict_false(err != 0)) { 1913 /* ARM64TODO: Add mbuf defragmenting if we lack maps */ 1914 m_freem(*mbufp); 1915 *mbufp = NULL; 1916 return (err); 1917 } 1918 1919 /* Set how many subdescriptors is required */ 1920 subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT + nsegs - 1; 1921 if (subdesc_cnt > sq->free_cnt) { 1922 /* ARM64TODO: Add mbuf defragmentation if we lack descriptors */ 1923 bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap); 1924 return (ENOBUFS); 1925 } 1926 1927 qentry = nicvf_get_sq_desc(sq, subdesc_cnt); 1928 1929 /* Add SQ header subdesc */ 1930 err = nicvf_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, *mbufp, 1931 (*mbufp)->m_pkthdr.len); 1932 if (err != 0) { 1933 nicvf_put_sq_desc(sq, subdesc_cnt); 1934 bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap); 1935 if (err == ENOBUFS) { 1936 m_freem(*mbufp); 1937 *mbufp = NULL; 1938 } 1939 return (err); 1940 } 1941 1942 /* Add SQ gather subdescs */ 1943 for (seg = 0; seg < nsegs; seg++) { 1944 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1945 nicvf_sq_add_gather_subdesc(sq, qentry, segs[seg].ds_len, 1946 segs[seg].ds_addr); 1947 } 1948 1949 /* make sure all memory stores are done before ringing doorbell */ 1950 bus_dmamap_sync(sq->dmem.dmat, sq->dmem.dmap, BUS_DMASYNC_PREWRITE); 1951 1952 dprintf(sq->nic->dev, "%s: sq->idx: %d, subdesc_cnt: %d\n", 1953 __func__, sq->idx, subdesc_cnt); 1954 /* Inform HW to xmit new packet */ 1955 nicvf_queue_reg_write(sq->nic, NIC_QSET_SQ_0_7_DOOR, 1956 sq->idx, subdesc_cnt); 1957 return (0); 1958 } 1959 1960 static __inline u_int 1961 frag_num(u_int i) 1962 { 1963 #if BYTE_ORDER == BIG_ENDIAN 1964 return ((i & ~3) + 3 - (i & 3)); 1965 #else 1966 return (i); 1967 #endif 1968 } 1969 1970 /* Returns MBUF for a received packet */ 1971 struct mbuf * 1972 nicvf_get_rcv_mbuf(struct nicvf *nic, struct cqe_rx_t *cqe_rx) 1973 { 1974 int frag; 1975 int payload_len = 0; 1976 struct mbuf *mbuf; 1977 struct mbuf *mbuf_frag; 1978 uint16_t *rb_lens = NULL; 1979 uint64_t *rb_ptrs = NULL; 1980 1981 mbuf = NULL; 1982 rb_lens = (uint16_t *)((uint8_t *)cqe_rx + (3 * sizeof(uint64_t))); 1983 rb_ptrs = (uint64_t *)((uint8_t *)cqe_rx + (6 * sizeof(uint64_t))); 1984 1985 dprintf(nic->dev, "%s rb_cnt %d rb0_ptr %lx rb0_sz %d\n", 1986 __func__, cqe_rx->rb_cnt, cqe_rx->rb0_ptr, cqe_rx->rb0_sz); 1987 1988 for (frag = 0; frag < cqe_rx->rb_cnt; frag++) { 1989 payload_len = rb_lens[frag_num(frag)]; 1990 if (frag == 0) { 1991 /* First fragment */ 1992 mbuf = nicvf_rb_ptr_to_mbuf(nic, 1993 (*rb_ptrs - cqe_rx->align_pad)); 1994 mbuf->m_len = payload_len; 1995 mbuf->m_data += cqe_rx->align_pad; 1996 if_setrcvif(mbuf, nic->ifp); 1997 } else { 1998 /* Add fragments */ 1999 mbuf_frag = nicvf_rb_ptr_to_mbuf(nic, *rb_ptrs); 2000 m_append(mbuf, payload_len, mbuf_frag->m_data); 2001 m_freem(mbuf_frag); 2002 } 2003 /* Next buffer pointer */ 2004 rb_ptrs++; 2005 } 2006 2007 if (__predict_true(mbuf != NULL)) { 2008 m_fixhdr(mbuf); 2009 mbuf->m_pkthdr.flowid = cqe_rx->rq_idx; 2010 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE); 2011 if (__predict_true((if_getcapenable(nic->ifp) & IFCAP_RXCSUM) != 0)) { 2012 /* 2013 * HW by default verifies IP & TCP/UDP/SCTP checksums 2014 */ 2015 if (__predict_true(cqe_rx->l3_type == L3TYPE_IPV4)) { 2016 mbuf->m_pkthdr.csum_flags = 2017 (CSUM_IP_CHECKED | CSUM_IP_VALID); 2018 } 2019 2020 switch (cqe_rx->l4_type) { 2021 case L4TYPE_UDP: 2022 case L4TYPE_TCP: /* fall through */ 2023 mbuf->m_pkthdr.csum_flags |= 2024 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 2025 mbuf->m_pkthdr.csum_data = 0xffff; 2026 break; 2027 case L4TYPE_SCTP: 2028 mbuf->m_pkthdr.csum_flags |= CSUM_SCTP_VALID; 2029 break; 2030 default: 2031 break; 2032 } 2033 } 2034 } 2035 2036 return (mbuf); 2037 } 2038 2039 /* Enable interrupt */ 2040 void 2041 nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx) 2042 { 2043 uint64_t reg_val; 2044 2045 reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S); 2046 2047 switch (int_type) { 2048 case NICVF_INTR_CQ: 2049 reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2050 break; 2051 case NICVF_INTR_SQ: 2052 reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2053 break; 2054 case NICVF_INTR_RBDR: 2055 reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2056 break; 2057 case NICVF_INTR_PKT_DROP: 2058 reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2059 break; 2060 case NICVF_INTR_TCP_TIMER: 2061 reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2062 break; 2063 case NICVF_INTR_MBOX: 2064 reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT); 2065 break; 2066 case NICVF_INTR_QS_ERR: 2067 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2068 break; 2069 default: 2070 device_printf(nic->dev, 2071 "Failed to enable interrupt: unknown type\n"); 2072 break; 2073 } 2074 2075 nicvf_reg_write(nic, NIC_VF_ENA_W1S, reg_val); 2076 } 2077 2078 /* Disable interrupt */ 2079 void 2080 nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx) 2081 { 2082 uint64_t reg_val = 0; 2083 2084 switch (int_type) { 2085 case NICVF_INTR_CQ: 2086 reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2087 break; 2088 case NICVF_INTR_SQ: 2089 reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2090 break; 2091 case NICVF_INTR_RBDR: 2092 reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2093 break; 2094 case NICVF_INTR_PKT_DROP: 2095 reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2096 break; 2097 case NICVF_INTR_TCP_TIMER: 2098 reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2099 break; 2100 case NICVF_INTR_MBOX: 2101 reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT); 2102 break; 2103 case NICVF_INTR_QS_ERR: 2104 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2105 break; 2106 default: 2107 device_printf(nic->dev, 2108 "Failed to disable interrupt: unknown type\n"); 2109 break; 2110 } 2111 2112 nicvf_reg_write(nic, NIC_VF_ENA_W1C, reg_val); 2113 } 2114 2115 /* Clear interrupt */ 2116 void 2117 nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx) 2118 { 2119 uint64_t reg_val = 0; 2120 2121 switch (int_type) { 2122 case NICVF_INTR_CQ: 2123 reg_val = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2124 break; 2125 case NICVF_INTR_SQ: 2126 reg_val = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2127 break; 2128 case NICVF_INTR_RBDR: 2129 reg_val = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2130 break; 2131 case NICVF_INTR_PKT_DROP: 2132 reg_val = (1UL << NICVF_INTR_PKT_DROP_SHIFT); 2133 break; 2134 case NICVF_INTR_TCP_TIMER: 2135 reg_val = (1UL << NICVF_INTR_TCP_TIMER_SHIFT); 2136 break; 2137 case NICVF_INTR_MBOX: 2138 reg_val = (1UL << NICVF_INTR_MBOX_SHIFT); 2139 break; 2140 case NICVF_INTR_QS_ERR: 2141 reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT); 2142 break; 2143 default: 2144 device_printf(nic->dev, 2145 "Failed to clear interrupt: unknown type\n"); 2146 break; 2147 } 2148 2149 nicvf_reg_write(nic, NIC_VF_INT, reg_val); 2150 } 2151 2152 /* Check if interrupt is enabled */ 2153 int 2154 nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx) 2155 { 2156 uint64_t reg_val; 2157 uint64_t mask = 0xff; 2158 2159 reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S); 2160 2161 switch (int_type) { 2162 case NICVF_INTR_CQ: 2163 mask = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT); 2164 break; 2165 case NICVF_INTR_SQ: 2166 mask = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT); 2167 break; 2168 case NICVF_INTR_RBDR: 2169 mask = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT); 2170 break; 2171 case NICVF_INTR_PKT_DROP: 2172 mask = NICVF_INTR_PKT_DROP_MASK; 2173 break; 2174 case NICVF_INTR_TCP_TIMER: 2175 mask = NICVF_INTR_TCP_TIMER_MASK; 2176 break; 2177 case NICVF_INTR_MBOX: 2178 mask = NICVF_INTR_MBOX_MASK; 2179 break; 2180 case NICVF_INTR_QS_ERR: 2181 mask = NICVF_INTR_QS_ERR_MASK; 2182 break; 2183 default: 2184 device_printf(nic->dev, 2185 "Failed to check interrupt enable: unknown type\n"); 2186 break; 2187 } 2188 2189 return (reg_val & mask); 2190 } 2191 2192 void 2193 nicvf_update_rq_stats(struct nicvf *nic, int rq_idx) 2194 { 2195 struct rcv_queue *rq; 2196 2197 #define GET_RQ_STATS(reg) \ 2198 nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\ 2199 (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 2200 2201 rq = &nic->qs->rq[rq_idx]; 2202 rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS); 2203 rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS); 2204 } 2205 2206 void 2207 nicvf_update_sq_stats(struct nicvf *nic, int sq_idx) 2208 { 2209 struct snd_queue *sq; 2210 2211 #define GET_SQ_STATS(reg) \ 2212 nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\ 2213 (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 2214 2215 sq = &nic->qs->sq[sq_idx]; 2216 sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS); 2217 sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS); 2218 } 2219 2220 /* Check for errors in the receive cmp.queue entry */ 2221 int 2222 nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cmp_queue *cq, 2223 struct cqe_rx_t *cqe_rx) 2224 { 2225 struct nicvf_hw_stats *stats = &nic->hw_stats; 2226 struct nicvf_drv_stats *drv_stats = &nic->drv_stats; 2227 2228 if (!cqe_rx->err_level && !cqe_rx->err_opcode) { 2229 drv_stats->rx_frames_ok++; 2230 return (0); 2231 } 2232 2233 switch (cqe_rx->err_opcode) { 2234 case CQ_RX_ERROP_RE_PARTIAL: 2235 stats->rx_bgx_truncated_pkts++; 2236 break; 2237 case CQ_RX_ERROP_RE_JABBER: 2238 stats->rx_jabber_errs++; 2239 break; 2240 case CQ_RX_ERROP_RE_FCS: 2241 stats->rx_fcs_errs++; 2242 break; 2243 case CQ_RX_ERROP_RE_RX_CTL: 2244 stats->rx_bgx_errs++; 2245 break; 2246 case CQ_RX_ERROP_PREL2_ERR: 2247 stats->rx_prel2_errs++; 2248 break; 2249 case CQ_RX_ERROP_L2_MAL: 2250 stats->rx_l2_hdr_malformed++; 2251 break; 2252 case CQ_RX_ERROP_L2_OVERSIZE: 2253 stats->rx_oversize++; 2254 break; 2255 case CQ_RX_ERROP_L2_UNDERSIZE: 2256 stats->rx_undersize++; 2257 break; 2258 case CQ_RX_ERROP_L2_LENMISM: 2259 stats->rx_l2_len_mismatch++; 2260 break; 2261 case CQ_RX_ERROP_L2_PCLP: 2262 stats->rx_l2_pclp++; 2263 break; 2264 case CQ_RX_ERROP_IP_NOT: 2265 stats->rx_ip_ver_errs++; 2266 break; 2267 case CQ_RX_ERROP_IP_CSUM_ERR: 2268 stats->rx_ip_csum_errs++; 2269 break; 2270 case CQ_RX_ERROP_IP_MAL: 2271 stats->rx_ip_hdr_malformed++; 2272 break; 2273 case CQ_RX_ERROP_IP_MALD: 2274 stats->rx_ip_payload_malformed++; 2275 break; 2276 case CQ_RX_ERROP_IP_HOP: 2277 stats->rx_ip_ttl_errs++; 2278 break; 2279 case CQ_RX_ERROP_L3_PCLP: 2280 stats->rx_l3_pclp++; 2281 break; 2282 case CQ_RX_ERROP_L4_MAL: 2283 stats->rx_l4_malformed++; 2284 break; 2285 case CQ_RX_ERROP_L4_CHK: 2286 stats->rx_l4_csum_errs++; 2287 break; 2288 case CQ_RX_ERROP_UDP_LEN: 2289 stats->rx_udp_len_errs++; 2290 break; 2291 case CQ_RX_ERROP_L4_PORT: 2292 stats->rx_l4_port_errs++; 2293 break; 2294 case CQ_RX_ERROP_TCP_FLAG: 2295 stats->rx_tcp_flag_errs++; 2296 break; 2297 case CQ_RX_ERROP_TCP_OFFSET: 2298 stats->rx_tcp_offset_errs++; 2299 break; 2300 case CQ_RX_ERROP_L4_PCLP: 2301 stats->rx_l4_pclp++; 2302 break; 2303 case CQ_RX_ERROP_RBDR_TRUNC: 2304 stats->rx_truncated_pkts++; 2305 break; 2306 } 2307 2308 return (1); 2309 } 2310 2311 /* Check for errors in the send cmp.queue entry */ 2312 int 2313 nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cmp_queue *cq, 2314 struct cqe_send_t *cqe_tx) 2315 { 2316 struct cmp_queue_stats *stats = &cq->stats; 2317 2318 switch (cqe_tx->send_status) { 2319 case CQ_TX_ERROP_GOOD: 2320 stats->tx.good++; 2321 return (0); 2322 case CQ_TX_ERROP_DESC_FAULT: 2323 stats->tx.desc_fault++; 2324 break; 2325 case CQ_TX_ERROP_HDR_CONS_ERR: 2326 stats->tx.hdr_cons_err++; 2327 break; 2328 case CQ_TX_ERROP_SUBDC_ERR: 2329 stats->tx.subdesc_err++; 2330 break; 2331 case CQ_TX_ERROP_IMM_SIZE_OFLOW: 2332 stats->tx.imm_size_oflow++; 2333 break; 2334 case CQ_TX_ERROP_DATA_SEQUENCE_ERR: 2335 stats->tx.data_seq_err++; 2336 break; 2337 case CQ_TX_ERROP_MEM_SEQUENCE_ERR: 2338 stats->tx.mem_seq_err++; 2339 break; 2340 case CQ_TX_ERROP_LOCK_VIOL: 2341 stats->tx.lock_viol++; 2342 break; 2343 case CQ_TX_ERROP_DATA_FAULT: 2344 stats->tx.data_fault++; 2345 break; 2346 case CQ_TX_ERROP_TSTMP_CONFLICT: 2347 stats->tx.tstmp_conflict++; 2348 break; 2349 case CQ_TX_ERROP_TSTMP_TIMEOUT: 2350 stats->tx.tstmp_timeout++; 2351 break; 2352 case CQ_TX_ERROP_MEM_FAULT: 2353 stats->tx.mem_fault++; 2354 break; 2355 case CQ_TX_ERROP_CK_OVERLAP: 2356 stats->tx.csum_overlap++; 2357 break; 2358 case CQ_TX_ERROP_CK_OFLOW: 2359 stats->tx.csum_overflow++; 2360 break; 2361 } 2362 2363 return (1); 2364 } 2365