1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2015 Cavium, Inc. 4 */ 5 6 #include <linux/pci.h> 7 #include <linux/netdevice.h> 8 #include <linux/ip.h> 9 #include <linux/etherdevice.h> 10 #include <linux/iommu.h> 11 #include <net/ip.h> 12 #include <net/tso.h> 13 14 #include "nic_reg.h" 15 #include "nic.h" 16 #include "q_struct.h" 17 #include "nicvf_queues.h" 18 19 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry, 20 int size, u64 data); 21 static void nicvf_get_page(struct nicvf *nic) 22 { 23 if (!nic->rb_pageref || !nic->rb_page) 24 return; 25 26 page_ref_add(nic->rb_page, nic->rb_pageref); 27 nic->rb_pageref = 0; 28 } 29 30 /* Poll a register for a specific value */ 31 static int nicvf_poll_reg(struct nicvf *nic, int qidx, 32 u64 reg, int bit_pos, int bits, int val) 33 { 34 u64 bit_mask; 35 u64 reg_val; 36 int timeout = 10; 37 38 bit_mask = (1ULL << bits) - 1; 39 bit_mask = (bit_mask << bit_pos); 40 41 while (timeout) { 42 reg_val = nicvf_queue_reg_read(nic, reg, qidx); 43 if (((reg_val & bit_mask) >> bit_pos) == val) 44 return 0; 45 usleep_range(1000, 2000); 46 timeout--; 47 } 48 netdev_err(nic->netdev, "Poll on reg 0x%llx failed\n", reg); 49 return 1; 50 } 51 52 /* Allocate memory for a queue's descriptors */ 53 static int nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem, 54 int q_len, int desc_size, int align_bytes) 55 { 56 dmem->q_len = q_len; 57 dmem->size = (desc_size * q_len) + align_bytes; 58 /* Save address, need it while freeing */ 59 dmem->unalign_base = dma_alloc_coherent(&nic->pdev->dev, dmem->size, 60 &dmem->dma, GFP_KERNEL); 61 if (!dmem->unalign_base) 62 return -ENOMEM; 63 64 /* Align memory address for 'align_bytes' */ 65 dmem->phys_base = NICVF_ALIGNED_ADDR((u64)dmem->dma, align_bytes); 66 dmem->base = dmem->unalign_base + (dmem->phys_base - dmem->dma); 67 return 0; 68 } 69 70 /* Free queue's descriptor memory */ 71 static void nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem) 72 { 73 if (!dmem) 74 return; 75 76 dma_free_coherent(&nic->pdev->dev, dmem->size, 77 dmem->unalign_base, dmem->dma); 78 dmem->unalign_base = NULL; 79 dmem->base = NULL; 80 } 81 82 #define XDP_PAGE_REFCNT_REFILL 256 83 84 /* Allocate a new page or recycle one if possible 85 * 86 * We cannot optimize dma mapping here, since 87 * 1. It's only one RBDR ring for 8 Rx queues. 88 * 2. CQE_RX gives address of the buffer where pkt has been DMA'ed 89 * and not idx into RBDR ring, so can't refer to saved info. 90 * 3. There are multiple receive buffers per page 91 */ 92 static inline struct pgcache *nicvf_alloc_page(struct nicvf *nic, 93 struct rbdr *rbdr, gfp_t gfp) 94 { 95 int ref_count; 96 struct page *page = NULL; 97 struct pgcache *pgcache, *next; 98 99 /* Check if page is already allocated */ 100 pgcache = &rbdr->pgcache[rbdr->pgidx]; 101 page = pgcache->page; 102 /* Check if page can be recycled */ 103 if (page) { 104 ref_count = page_ref_count(page); 105 /* This page can be recycled if internal ref_count and page's 106 * ref_count are equal, indicating that the page has been used 107 * once for packet transmission. For non-XDP mode, internal 108 * ref_count is always '1'. 109 */ 110 if (rbdr->is_xdp) { 111 if (ref_count == pgcache->ref_count) 112 pgcache->ref_count--; 113 else 114 page = NULL; 115 } else if (ref_count != 1) { 116 page = NULL; 117 } 118 } 119 120 if (!page) { 121 page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN, 0); 122 if (!page) 123 return NULL; 124 125 this_cpu_inc(nic->pnicvf->drv_stats->page_alloc); 126 127 /* Check for space */ 128 if (rbdr->pgalloc >= rbdr->pgcnt) { 129 /* Page can still be used */ 130 nic->rb_page = page; 131 return NULL; 132 } 133 134 /* Save the page in page cache */ 135 pgcache->page = page; 136 pgcache->dma_addr = 0; 137 pgcache->ref_count = 0; 138 rbdr->pgalloc++; 139 } 140 141 /* Take additional page references for recycling */ 142 if (rbdr->is_xdp) { 143 /* Since there is single RBDR (i.e single core doing 144 * page recycling) per 8 Rx queues, in XDP mode adjusting 145 * page references atomically is the biggest bottleneck, so 146 * take bunch of references at a time. 147 * 148 * So here, below reference counts defer by '1'. 149 */ 150 if (!pgcache->ref_count) { 151 pgcache->ref_count = XDP_PAGE_REFCNT_REFILL; 152 page_ref_add(page, XDP_PAGE_REFCNT_REFILL); 153 } 154 } else { 155 /* In non-XDP case, single 64K page is divided across multiple 156 * receive buffers, so cost of recycling is less anyway. 157 * So we can do with just one extra reference. 158 */ 159 page_ref_add(page, 1); 160 } 161 162 rbdr->pgidx++; 163 rbdr->pgidx &= (rbdr->pgcnt - 1); 164 165 /* Prefetch refcount of next page in page cache */ 166 next = &rbdr->pgcache[rbdr->pgidx]; 167 page = next->page; 168 if (page) 169 prefetch(&page->_refcount); 170 171 return pgcache; 172 } 173 174 /* Allocate buffer for packet reception */ 175 static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr, 176 gfp_t gfp, u32 buf_len, u64 *rbuf) 177 { 178 struct pgcache *pgcache = NULL; 179 180 /* Check if request can be accomodated in previous allocated page. 181 * But in XDP mode only one buffer per page is permitted. 182 */ 183 if (!rbdr->is_xdp && nic->rb_page && 184 ((nic->rb_page_offset + buf_len) <= PAGE_SIZE)) { 185 nic->rb_pageref++; 186 goto ret; 187 } 188 189 nicvf_get_page(nic); 190 nic->rb_page = NULL; 191 192 /* Get new page, either recycled or new one */ 193 pgcache = nicvf_alloc_page(nic, rbdr, gfp); 194 if (!pgcache && !nic->rb_page) { 195 this_cpu_inc(nic->pnicvf->drv_stats->rcv_buffer_alloc_failures); 196 return -ENOMEM; 197 } 198 199 nic->rb_page_offset = 0; 200 201 /* Reserve space for header modifications by BPF program */ 202 if (rbdr->is_xdp) 203 buf_len += XDP_PACKET_HEADROOM; 204 205 /* Check if it's recycled */ 206 if (pgcache) 207 nic->rb_page = pgcache->page; 208 ret: 209 if (rbdr->is_xdp && pgcache && pgcache->dma_addr) { 210 *rbuf = pgcache->dma_addr; 211 } else { 212 /* HW will ensure data coherency, CPU sync not required */ 213 *rbuf = (u64)dma_map_page_attrs(&nic->pdev->dev, nic->rb_page, 214 nic->rb_page_offset, buf_len, 215 DMA_FROM_DEVICE, 216 DMA_ATTR_SKIP_CPU_SYNC); 217 if (dma_mapping_error(&nic->pdev->dev, (dma_addr_t)*rbuf)) { 218 if (!nic->rb_page_offset) 219 __free_pages(nic->rb_page, 0); 220 nic->rb_page = NULL; 221 return -ENOMEM; 222 } 223 if (pgcache) 224 pgcache->dma_addr = *rbuf + XDP_PACKET_HEADROOM; 225 nic->rb_page_offset += buf_len; 226 } 227 228 return 0; 229 } 230 231 /* Build skb around receive buffer */ 232 static struct sk_buff *nicvf_rb_ptr_to_skb(struct nicvf *nic, 233 u64 rb_ptr, int len) 234 { 235 void *data; 236 struct sk_buff *skb; 237 238 data = phys_to_virt(rb_ptr); 239 240 /* Now build an skb to give to stack */ 241 skb = build_skb(data, RCV_FRAG_LEN); 242 if (!skb) { 243 put_page(virt_to_page(data)); 244 return NULL; 245 } 246 247 prefetch(skb->data); 248 return skb; 249 } 250 251 /* Allocate RBDR ring and populate receive buffers */ 252 static int nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr, 253 int ring_len, int buf_size) 254 { 255 int idx; 256 u64 rbuf; 257 struct rbdr_entry_t *desc; 258 int err; 259 260 err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len, 261 sizeof(struct rbdr_entry_t), 262 NICVF_RCV_BUF_ALIGN_BYTES); 263 if (err) 264 return err; 265 266 rbdr->desc = rbdr->dmem.base; 267 /* Buffer size has to be in multiples of 128 bytes */ 268 rbdr->dma_size = buf_size; 269 rbdr->enable = true; 270 rbdr->thresh = RBDR_THRESH; 271 rbdr->head = 0; 272 rbdr->tail = 0; 273 274 /* Initialize page recycling stuff. 275 * 276 * Can't use single buffer per page especially with 64K pages. 277 * On embedded platforms i.e 81xx/83xx available memory itself 278 * is low and minimum ring size of RBDR is 8K, that takes away 279 * lots of memory. 280 * 281 * But for XDP it has to be a single buffer per page. 282 */ 283 if (!nic->pnicvf->xdp_prog) { 284 rbdr->pgcnt = ring_len / (PAGE_SIZE / buf_size); 285 rbdr->is_xdp = false; 286 } else { 287 rbdr->pgcnt = ring_len; 288 rbdr->is_xdp = true; 289 } 290 rbdr->pgcnt = roundup_pow_of_two(rbdr->pgcnt); 291 rbdr->pgcache = kcalloc(rbdr->pgcnt, sizeof(*rbdr->pgcache), 292 GFP_KERNEL); 293 if (!rbdr->pgcache) 294 return -ENOMEM; 295 rbdr->pgidx = 0; 296 rbdr->pgalloc = 0; 297 298 nic->rb_page = NULL; 299 for (idx = 0; idx < ring_len; idx++) { 300 err = nicvf_alloc_rcv_buffer(nic, rbdr, GFP_KERNEL, 301 RCV_FRAG_LEN, &rbuf); 302 if (err) { 303 /* To free already allocated and mapped ones */ 304 rbdr->tail = idx - 1; 305 return err; 306 } 307 308 desc = GET_RBDR_DESC(rbdr, idx); 309 desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1); 310 } 311 312 nicvf_get_page(nic); 313 314 return 0; 315 } 316 317 /* Free RBDR ring and its receive buffers */ 318 static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr) 319 { 320 int head, tail; 321 u64 buf_addr, phys_addr; 322 struct pgcache *pgcache; 323 struct rbdr_entry_t *desc; 324 325 if (!rbdr) 326 return; 327 328 rbdr->enable = false; 329 if (!rbdr->dmem.base) 330 return; 331 332 head = rbdr->head; 333 tail = rbdr->tail; 334 335 /* Release page references */ 336 while (head != tail) { 337 desc = GET_RBDR_DESC(rbdr, head); 338 buf_addr = desc->buf_addr; 339 phys_addr = nicvf_iova_to_phys(nic, buf_addr); 340 dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN, 341 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC); 342 if (phys_addr) 343 put_page(virt_to_page(phys_to_virt(phys_addr))); 344 head++; 345 head &= (rbdr->dmem.q_len - 1); 346 } 347 /* Release buffer of tail desc */ 348 desc = GET_RBDR_DESC(rbdr, tail); 349 buf_addr = desc->buf_addr; 350 phys_addr = nicvf_iova_to_phys(nic, buf_addr); 351 dma_unmap_page_attrs(&nic->pdev->dev, buf_addr, RCV_FRAG_LEN, 352 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC); 353 if (phys_addr) 354 put_page(virt_to_page(phys_to_virt(phys_addr))); 355 356 /* Sync page cache info */ 357 smp_rmb(); 358 359 /* Release additional page references held for recycling */ 360 head = 0; 361 while (head < rbdr->pgcnt) { 362 pgcache = &rbdr->pgcache[head]; 363 if (pgcache->page && page_ref_count(pgcache->page) != 0) { 364 if (rbdr->is_xdp) { 365 page_ref_sub(pgcache->page, 366 pgcache->ref_count - 1); 367 } 368 put_page(pgcache->page); 369 } 370 head++; 371 } 372 373 /* Free RBDR ring */ 374 nicvf_free_q_desc_mem(nic, &rbdr->dmem); 375 } 376 377 /* Refill receive buffer descriptors with new buffers. 378 */ 379 static void nicvf_refill_rbdr(struct nicvf *nic, gfp_t gfp) 380 { 381 struct queue_set *qs = nic->qs; 382 int rbdr_idx = qs->rbdr_cnt; 383 int tail, qcount; 384 int refill_rb_cnt; 385 struct rbdr *rbdr; 386 struct rbdr_entry_t *desc; 387 u64 rbuf; 388 int new_rb = 0; 389 390 refill: 391 if (!rbdr_idx) 392 return; 393 rbdr_idx--; 394 rbdr = &qs->rbdr[rbdr_idx]; 395 /* Check if it's enabled */ 396 if (!rbdr->enable) 397 goto next_rbdr; 398 399 /* Get no of desc's to be refilled */ 400 qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx); 401 qcount &= 0x7FFFF; 402 /* Doorbell can be ringed with a max of ring size minus 1 */ 403 if (qcount >= (qs->rbdr_len - 1)) 404 goto next_rbdr; 405 else 406 refill_rb_cnt = qs->rbdr_len - qcount - 1; 407 408 /* Sync page cache info */ 409 smp_rmb(); 410 411 /* Start filling descs from tail */ 412 tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3; 413 while (refill_rb_cnt) { 414 tail++; 415 tail &= (rbdr->dmem.q_len - 1); 416 417 if (nicvf_alloc_rcv_buffer(nic, rbdr, gfp, RCV_FRAG_LEN, &rbuf)) 418 break; 419 420 desc = GET_RBDR_DESC(rbdr, tail); 421 desc->buf_addr = rbuf & ~(NICVF_RCV_BUF_ALIGN_BYTES - 1); 422 refill_rb_cnt--; 423 new_rb++; 424 } 425 426 nicvf_get_page(nic); 427 428 /* make sure all memory stores are done before ringing doorbell */ 429 smp_wmb(); 430 431 /* Check if buffer allocation failed */ 432 if (refill_rb_cnt) 433 nic->rb_alloc_fail = true; 434 else 435 nic->rb_alloc_fail = false; 436 437 /* Notify HW */ 438 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, 439 rbdr_idx, new_rb); 440 next_rbdr: 441 /* Re-enable RBDR interrupts only if buffer allocation is success */ 442 if (!nic->rb_alloc_fail && rbdr->enable && 443 netif_running(nic->pnicvf->netdev)) 444 nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx); 445 446 if (rbdr_idx) 447 goto refill; 448 } 449 450 /* Alloc rcv buffers in non-atomic mode for better success */ 451 void nicvf_rbdr_work(struct work_struct *work) 452 { 453 struct nicvf *nic = container_of(work, struct nicvf, rbdr_work.work); 454 455 nicvf_refill_rbdr(nic, GFP_KERNEL); 456 if (nic->rb_alloc_fail) 457 schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10)); 458 else 459 nic->rb_work_scheduled = false; 460 } 461 462 /* In Softirq context, alloc rcv buffers in atomic mode */ 463 void nicvf_rbdr_task(unsigned long data) 464 { 465 struct nicvf *nic = (struct nicvf *)data; 466 467 nicvf_refill_rbdr(nic, GFP_ATOMIC); 468 if (nic->rb_alloc_fail) { 469 nic->rb_work_scheduled = true; 470 schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10)); 471 } 472 } 473 474 /* Initialize completion queue */ 475 static int nicvf_init_cmp_queue(struct nicvf *nic, 476 struct cmp_queue *cq, int q_len) 477 { 478 int err; 479 480 err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE, 481 NICVF_CQ_BASE_ALIGN_BYTES); 482 if (err) 483 return err; 484 485 cq->desc = cq->dmem.base; 486 cq->thresh = pass1_silicon(nic->pdev) ? 0 : CMP_QUEUE_CQE_THRESH; 487 nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1; 488 489 return 0; 490 } 491 492 static void nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq) 493 { 494 if (!cq) 495 return; 496 if (!cq->dmem.base) 497 return; 498 499 nicvf_free_q_desc_mem(nic, &cq->dmem); 500 } 501 502 /* Initialize transmit queue */ 503 static int nicvf_init_snd_queue(struct nicvf *nic, 504 struct snd_queue *sq, int q_len, int qidx) 505 { 506 int err; 507 508 err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE, 509 NICVF_SQ_BASE_ALIGN_BYTES); 510 if (err) 511 return err; 512 513 sq->desc = sq->dmem.base; 514 sq->skbuff = kcalloc(q_len, sizeof(u64), GFP_KERNEL); 515 if (!sq->skbuff) 516 return -ENOMEM; 517 518 sq->head = 0; 519 sq->tail = 0; 520 sq->thresh = SND_QUEUE_THRESH; 521 522 /* Check if this SQ is a XDP TX queue */ 523 if (nic->sqs_mode) 524 qidx += ((nic->sqs_id + 1) * MAX_SND_QUEUES_PER_QS); 525 if (qidx < nic->pnicvf->xdp_tx_queues) { 526 /* Alloc memory to save page pointers for XDP_TX */ 527 sq->xdp_page = kcalloc(q_len, sizeof(u64), GFP_KERNEL); 528 if (!sq->xdp_page) 529 return -ENOMEM; 530 sq->xdp_desc_cnt = 0; 531 sq->xdp_free_cnt = q_len - 1; 532 sq->is_xdp = true; 533 } else { 534 sq->xdp_page = NULL; 535 sq->xdp_desc_cnt = 0; 536 sq->xdp_free_cnt = 0; 537 sq->is_xdp = false; 538 539 atomic_set(&sq->free_cnt, q_len - 1); 540 541 /* Preallocate memory for TSO segment's header */ 542 sq->tso_hdrs = dma_alloc_coherent(&nic->pdev->dev, 543 q_len * TSO_HEADER_SIZE, 544 &sq->tso_hdrs_phys, 545 GFP_KERNEL); 546 if (!sq->tso_hdrs) 547 return -ENOMEM; 548 } 549 550 return 0; 551 } 552 553 void nicvf_unmap_sndq_buffers(struct nicvf *nic, struct snd_queue *sq, 554 int hdr_sqe, u8 subdesc_cnt) 555 { 556 u8 idx; 557 struct sq_gather_subdesc *gather; 558 559 /* Unmap DMA mapped skb data buffers */ 560 for (idx = 0; idx < subdesc_cnt; idx++) { 561 hdr_sqe++; 562 hdr_sqe &= (sq->dmem.q_len - 1); 563 gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, hdr_sqe); 564 /* HW will ensure data coherency, CPU sync not required */ 565 dma_unmap_page_attrs(&nic->pdev->dev, gather->addr, 566 gather->size, DMA_TO_DEVICE, 567 DMA_ATTR_SKIP_CPU_SYNC); 568 } 569 } 570 571 static void nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq) 572 { 573 struct sk_buff *skb; 574 struct page *page; 575 struct sq_hdr_subdesc *hdr; 576 struct sq_hdr_subdesc *tso_sqe; 577 578 if (!sq) 579 return; 580 if (!sq->dmem.base) 581 return; 582 583 if (sq->tso_hdrs) { 584 dma_free_coherent(&nic->pdev->dev, 585 sq->dmem.q_len * TSO_HEADER_SIZE, 586 sq->tso_hdrs, sq->tso_hdrs_phys); 587 sq->tso_hdrs = NULL; 588 } 589 590 /* Free pending skbs in the queue */ 591 smp_rmb(); 592 while (sq->head != sq->tail) { 593 skb = (struct sk_buff *)sq->skbuff[sq->head]; 594 if (!skb || !sq->xdp_page) 595 goto next; 596 597 page = (struct page *)sq->xdp_page[sq->head]; 598 if (!page) 599 goto next; 600 else 601 put_page(page); 602 603 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head); 604 /* Check for dummy descriptor used for HW TSO offload on 88xx */ 605 if (hdr->dont_send) { 606 /* Get actual TSO descriptors and unmap them */ 607 tso_sqe = 608 (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, hdr->rsvd2); 609 nicvf_unmap_sndq_buffers(nic, sq, hdr->rsvd2, 610 tso_sqe->subdesc_cnt); 611 } else { 612 nicvf_unmap_sndq_buffers(nic, sq, sq->head, 613 hdr->subdesc_cnt); 614 } 615 if (skb) 616 dev_kfree_skb_any(skb); 617 next: 618 sq->head++; 619 sq->head &= (sq->dmem.q_len - 1); 620 } 621 kfree(sq->skbuff); 622 kfree(sq->xdp_page); 623 nicvf_free_q_desc_mem(nic, &sq->dmem); 624 } 625 626 static void nicvf_reclaim_snd_queue(struct nicvf *nic, 627 struct queue_set *qs, int qidx) 628 { 629 /* Disable send queue */ 630 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0); 631 /* Check if SQ is stopped */ 632 if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01)) 633 return; 634 /* Reset send queue */ 635 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 636 } 637 638 static void nicvf_reclaim_rcv_queue(struct nicvf *nic, 639 struct queue_set *qs, int qidx) 640 { 641 union nic_mbx mbx = {}; 642 643 /* Make sure all packets in the pipeline are written back into mem */ 644 mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC; 645 nicvf_send_msg_to_pf(nic, &mbx); 646 } 647 648 static void nicvf_reclaim_cmp_queue(struct nicvf *nic, 649 struct queue_set *qs, int qidx) 650 { 651 /* Disable timer threshold (doesn't get reset upon CQ reset */ 652 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0); 653 /* Disable completion queue */ 654 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0); 655 /* Reset completion queue */ 656 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 657 } 658 659 static void nicvf_reclaim_rbdr(struct nicvf *nic, 660 struct rbdr *rbdr, int qidx) 661 { 662 u64 tmp, fifo_state; 663 int timeout = 10; 664 665 /* Save head and tail pointers for feeing up buffers */ 666 rbdr->head = nicvf_queue_reg_read(nic, 667 NIC_QSET_RBDR_0_1_HEAD, 668 qidx) >> 3; 669 rbdr->tail = nicvf_queue_reg_read(nic, 670 NIC_QSET_RBDR_0_1_TAIL, 671 qidx) >> 3; 672 673 /* If RBDR FIFO is in 'FAIL' state then do a reset first 674 * before relaiming. 675 */ 676 fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx); 677 if (((fifo_state >> 62) & 0x03) == 0x3) 678 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, 679 qidx, NICVF_RBDR_RESET); 680 681 /* Disable RBDR */ 682 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0); 683 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 684 return; 685 while (1) { 686 tmp = nicvf_queue_reg_read(nic, 687 NIC_QSET_RBDR_0_1_PREFETCH_STATUS, 688 qidx); 689 if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF)) 690 break; 691 usleep_range(1000, 2000); 692 timeout--; 693 if (!timeout) { 694 netdev_err(nic->netdev, 695 "Failed polling on prefetch status\n"); 696 return; 697 } 698 } 699 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, 700 qidx, NICVF_RBDR_RESET); 701 702 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02)) 703 return; 704 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00); 705 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00)) 706 return; 707 } 708 709 void nicvf_config_vlan_stripping(struct nicvf *nic, netdev_features_t features) 710 { 711 u64 rq_cfg; 712 int sqs; 713 714 rq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_RQ_GEN_CFG, 0); 715 716 /* Enable first VLAN stripping */ 717 if (features & NETIF_F_HW_VLAN_CTAG_RX) 718 rq_cfg |= (1ULL << 25); 719 else 720 rq_cfg &= ~(1ULL << 25); 721 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, rq_cfg); 722 723 /* Configure Secondary Qsets, if any */ 724 for (sqs = 0; sqs < nic->sqs_count; sqs++) 725 if (nic->snicvf[sqs]) 726 nicvf_queue_reg_write(nic->snicvf[sqs], 727 NIC_QSET_RQ_GEN_CFG, 0, rq_cfg); 728 } 729 730 static void nicvf_reset_rcv_queue_stats(struct nicvf *nic) 731 { 732 union nic_mbx mbx = {}; 733 734 /* Reset all RQ/SQ and VF stats */ 735 mbx.reset_stat.msg = NIC_MBOX_MSG_RESET_STAT_COUNTER; 736 mbx.reset_stat.rx_stat_mask = 0x3FFF; 737 mbx.reset_stat.tx_stat_mask = 0x1F; 738 mbx.reset_stat.rq_stat_mask = 0xFFFF; 739 mbx.reset_stat.sq_stat_mask = 0xFFFF; 740 nicvf_send_msg_to_pf(nic, &mbx); 741 } 742 743 /* Configures receive queue */ 744 static void nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs, 745 int qidx, bool enable) 746 { 747 union nic_mbx mbx = {}; 748 struct rcv_queue *rq; 749 struct rq_cfg rq_cfg; 750 751 rq = &qs->rq[qidx]; 752 rq->enable = enable; 753 754 /* Disable receive queue */ 755 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0); 756 757 if (!rq->enable) { 758 nicvf_reclaim_rcv_queue(nic, qs, qidx); 759 xdp_rxq_info_unreg(&rq->xdp_rxq); 760 return; 761 } 762 763 rq->cq_qs = qs->vnic_id; 764 rq->cq_idx = qidx; 765 rq->start_rbdr_qs = qs->vnic_id; 766 rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1; 767 rq->cont_rbdr_qs = qs->vnic_id; 768 rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1; 769 /* all writes of RBDR data to be loaded into L2 Cache as well*/ 770 rq->caching = 1; 771 772 /* Driver have no proper error path for failed XDP RX-queue info reg */ 773 WARN_ON(xdp_rxq_info_reg(&rq->xdp_rxq, nic->netdev, qidx) < 0); 774 775 /* Send a mailbox msg to PF to config RQ */ 776 mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG; 777 mbx.rq.qs_num = qs->vnic_id; 778 mbx.rq.rq_num = qidx; 779 mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) | 780 (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) | 781 (rq->cont_qs_rbdr_idx << 8) | 782 (rq->start_rbdr_qs << 1) | (rq->start_qs_rbdr_idx); 783 nicvf_send_msg_to_pf(nic, &mbx); 784 785 mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG; 786 mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) | 787 (RQ_PASS_RBDR_LVL << 16) | (RQ_PASS_CQ_LVL << 8) | 788 (qs->vnic_id << 0); 789 nicvf_send_msg_to_pf(nic, &mbx); 790 791 /* RQ drop config 792 * Enable CQ drop to reserve sufficient CQEs for all tx packets 793 */ 794 mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG; 795 mbx.rq.cfg = BIT_ULL(63) | BIT_ULL(62) | 796 (RQ_PASS_RBDR_LVL << 40) | (RQ_DROP_RBDR_LVL << 32) | 797 (RQ_PASS_CQ_LVL << 16) | (RQ_DROP_CQ_LVL << 8); 798 nicvf_send_msg_to_pf(nic, &mbx); 799 800 if (!nic->sqs_mode && (qidx == 0)) { 801 /* Enable checking L3/L4 length and TCP/UDP checksums 802 * Also allow IPv6 pkts with zero UDP checksum. 803 */ 804 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, 805 (BIT(24) | BIT(23) | BIT(21) | BIT(20))); 806 nicvf_config_vlan_stripping(nic, nic->netdev->features); 807 } 808 809 /* Enable Receive queue */ 810 memset(&rq_cfg, 0, sizeof(struct rq_cfg)); 811 rq_cfg.ena = 1; 812 rq_cfg.tcp_ena = 0; 813 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, *(u64 *)&rq_cfg); 814 } 815 816 /* Configures completion queue */ 817 void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs, 818 int qidx, bool enable) 819 { 820 struct cmp_queue *cq; 821 struct cq_cfg cq_cfg; 822 823 cq = &qs->cq[qidx]; 824 cq->enable = enable; 825 826 if (!cq->enable) { 827 nicvf_reclaim_cmp_queue(nic, qs, qidx); 828 return; 829 } 830 831 /* Reset completion queue */ 832 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET); 833 834 if (!cq->enable) 835 return; 836 837 spin_lock_init(&cq->lock); 838 /* Set completion queue base address */ 839 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE, 840 qidx, (u64)(cq->dmem.phys_base)); 841 842 /* Enable Completion queue */ 843 memset(&cq_cfg, 0, sizeof(struct cq_cfg)); 844 cq_cfg.ena = 1; 845 cq_cfg.reset = 0; 846 cq_cfg.caching = 0; 847 cq_cfg.qsize = ilog2(qs->cq_len >> 10); 848 cq_cfg.avg_con = 0; 849 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg); 850 851 /* Set threshold value for interrupt generation */ 852 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh); 853 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, 854 qidx, CMP_QUEUE_TIMER_THRESH); 855 } 856 857 /* Configures transmit queue */ 858 static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs, 859 int qidx, bool enable) 860 { 861 union nic_mbx mbx = {}; 862 struct snd_queue *sq; 863 struct sq_cfg sq_cfg; 864 865 sq = &qs->sq[qidx]; 866 sq->enable = enable; 867 868 if (!sq->enable) { 869 nicvf_reclaim_snd_queue(nic, qs, qidx); 870 return; 871 } 872 873 /* Reset send queue */ 874 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET); 875 876 sq->cq_qs = qs->vnic_id; 877 sq->cq_idx = qidx; 878 879 /* Send a mailbox msg to PF to config SQ */ 880 mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG; 881 mbx.sq.qs_num = qs->vnic_id; 882 mbx.sq.sq_num = qidx; 883 mbx.sq.sqs_mode = nic->sqs_mode; 884 mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx; 885 nicvf_send_msg_to_pf(nic, &mbx); 886 887 /* Set queue base address */ 888 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE, 889 qidx, (u64)(sq->dmem.phys_base)); 890 891 /* Enable send queue & set queue size */ 892 memset(&sq_cfg, 0, sizeof(struct sq_cfg)); 893 sq_cfg.ena = 1; 894 sq_cfg.reset = 0; 895 sq_cfg.ldwb = 0; 896 sq_cfg.qsize = ilog2(qs->sq_len >> 10); 897 sq_cfg.tstmp_bgx_intf = 0; 898 /* CQ's level at which HW will stop processing SQEs to avoid 899 * transmitting a pkt with no space in CQ to post CQE_TX. 900 */ 901 sq_cfg.cq_limit = (CMP_QUEUE_PIPELINE_RSVD * 256) / qs->cq_len; 902 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg); 903 904 /* Set threshold value for interrupt generation */ 905 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh); 906 907 /* Set queue:cpu affinity for better load distribution */ 908 if (cpu_online(qidx)) { 909 cpumask_set_cpu(qidx, &sq->affinity_mask); 910 netif_set_xps_queue(nic->netdev, 911 &sq->affinity_mask, qidx); 912 } 913 } 914 915 /* Configures receive buffer descriptor ring */ 916 static void nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs, 917 int qidx, bool enable) 918 { 919 struct rbdr *rbdr; 920 struct rbdr_cfg rbdr_cfg; 921 922 rbdr = &qs->rbdr[qidx]; 923 nicvf_reclaim_rbdr(nic, rbdr, qidx); 924 if (!enable) 925 return; 926 927 /* Set descriptor base address */ 928 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE, 929 qidx, (u64)(rbdr->dmem.phys_base)); 930 931 /* Enable RBDR & set queue size */ 932 /* Buffer size should be in multiples of 128 bytes */ 933 memset(&rbdr_cfg, 0, sizeof(struct rbdr_cfg)); 934 rbdr_cfg.ena = 1; 935 rbdr_cfg.reset = 0; 936 rbdr_cfg.ldwb = 0; 937 rbdr_cfg.qsize = RBDR_SIZE; 938 rbdr_cfg.avg_con = 0; 939 rbdr_cfg.lines = rbdr->dma_size / 128; 940 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, 941 qidx, *(u64 *)&rbdr_cfg); 942 943 /* Notify HW */ 944 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, 945 qidx, qs->rbdr_len - 1); 946 947 /* Set threshold value for interrupt generation */ 948 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH, 949 qidx, rbdr->thresh - 1); 950 } 951 952 /* Requests PF to assign and enable Qset */ 953 void nicvf_qset_config(struct nicvf *nic, bool enable) 954 { 955 union nic_mbx mbx = {}; 956 struct queue_set *qs = nic->qs; 957 struct qs_cfg *qs_cfg; 958 959 if (!qs) { 960 netdev_warn(nic->netdev, 961 "Qset is still not allocated, don't init queues\n"); 962 return; 963 } 964 965 qs->enable = enable; 966 qs->vnic_id = nic->vf_id; 967 968 /* Send a mailbox msg to PF to config Qset */ 969 mbx.qs.msg = NIC_MBOX_MSG_QS_CFG; 970 mbx.qs.num = qs->vnic_id; 971 mbx.qs.sqs_count = nic->sqs_count; 972 973 mbx.qs.cfg = 0; 974 qs_cfg = (struct qs_cfg *)&mbx.qs.cfg; 975 if (qs->enable) { 976 qs_cfg->ena = 1; 977 #ifdef __BIG_ENDIAN 978 qs_cfg->be = 1; 979 #endif 980 qs_cfg->vnic = qs->vnic_id; 981 /* Enable Tx timestamping capability */ 982 if (nic->ptp_clock) 983 qs_cfg->send_tstmp_ena = 1; 984 } 985 nicvf_send_msg_to_pf(nic, &mbx); 986 } 987 988 static void nicvf_free_resources(struct nicvf *nic) 989 { 990 int qidx; 991 struct queue_set *qs = nic->qs; 992 993 /* Free receive buffer descriptor ring */ 994 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 995 nicvf_free_rbdr(nic, &qs->rbdr[qidx]); 996 997 /* Free completion queue */ 998 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 999 nicvf_free_cmp_queue(nic, &qs->cq[qidx]); 1000 1001 /* Free send queue */ 1002 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1003 nicvf_free_snd_queue(nic, &qs->sq[qidx]); 1004 } 1005 1006 static int nicvf_alloc_resources(struct nicvf *nic) 1007 { 1008 int qidx; 1009 struct queue_set *qs = nic->qs; 1010 1011 /* Alloc receive buffer descriptor ring */ 1012 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) { 1013 if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len, 1014 DMA_BUFFER_LEN)) 1015 goto alloc_fail; 1016 } 1017 1018 /* Alloc send queue */ 1019 for (qidx = 0; qidx < qs->sq_cnt; qidx++) { 1020 if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx)) 1021 goto alloc_fail; 1022 } 1023 1024 /* Alloc completion queue */ 1025 for (qidx = 0; qidx < qs->cq_cnt; qidx++) { 1026 if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len)) 1027 goto alloc_fail; 1028 } 1029 1030 return 0; 1031 alloc_fail: 1032 nicvf_free_resources(nic); 1033 return -ENOMEM; 1034 } 1035 1036 int nicvf_set_qset_resources(struct nicvf *nic) 1037 { 1038 struct queue_set *qs; 1039 1040 qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL); 1041 if (!qs) 1042 return -ENOMEM; 1043 nic->qs = qs; 1044 1045 /* Set count of each queue */ 1046 qs->rbdr_cnt = DEFAULT_RBDR_CNT; 1047 qs->rq_cnt = min_t(u8, MAX_RCV_QUEUES_PER_QS, num_online_cpus()); 1048 qs->sq_cnt = min_t(u8, MAX_SND_QUEUES_PER_QS, num_online_cpus()); 1049 qs->cq_cnt = max_t(u8, qs->rq_cnt, qs->sq_cnt); 1050 1051 /* Set queue lengths */ 1052 qs->rbdr_len = RCV_BUF_COUNT; 1053 qs->sq_len = SND_QUEUE_LEN; 1054 qs->cq_len = CMP_QUEUE_LEN; 1055 1056 nic->rx_queues = qs->rq_cnt; 1057 nic->tx_queues = qs->sq_cnt; 1058 nic->xdp_tx_queues = 0; 1059 1060 return 0; 1061 } 1062 1063 int nicvf_config_data_transfer(struct nicvf *nic, bool enable) 1064 { 1065 bool disable = false; 1066 struct queue_set *qs = nic->qs; 1067 struct queue_set *pqs = nic->pnicvf->qs; 1068 int qidx; 1069 1070 if (!qs) 1071 return 0; 1072 1073 /* Take primary VF's queue lengths. 1074 * This is needed to take queue lengths set from ethtool 1075 * into consideration. 1076 */ 1077 if (nic->sqs_mode && pqs) { 1078 qs->cq_len = pqs->cq_len; 1079 qs->sq_len = pqs->sq_len; 1080 } 1081 1082 if (enable) { 1083 if (nicvf_alloc_resources(nic)) 1084 return -ENOMEM; 1085 1086 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1087 nicvf_snd_queue_config(nic, qs, qidx, enable); 1088 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1089 nicvf_cmp_queue_config(nic, qs, qidx, enable); 1090 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1091 nicvf_rbdr_config(nic, qs, qidx, enable); 1092 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1093 nicvf_rcv_queue_config(nic, qs, qidx, enable); 1094 } else { 1095 for (qidx = 0; qidx < qs->rq_cnt; qidx++) 1096 nicvf_rcv_queue_config(nic, qs, qidx, disable); 1097 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) 1098 nicvf_rbdr_config(nic, qs, qidx, disable); 1099 for (qidx = 0; qidx < qs->sq_cnt; qidx++) 1100 nicvf_snd_queue_config(nic, qs, qidx, disable); 1101 for (qidx = 0; qidx < qs->cq_cnt; qidx++) 1102 nicvf_cmp_queue_config(nic, qs, qidx, disable); 1103 1104 nicvf_free_resources(nic); 1105 } 1106 1107 /* Reset RXQ's stats. 1108 * SQ's stats will get reset automatically once SQ is reset. 1109 */ 1110 nicvf_reset_rcv_queue_stats(nic); 1111 1112 return 0; 1113 } 1114 1115 /* Get a free desc from SQ 1116 * returns descriptor ponter & descriptor number 1117 */ 1118 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt) 1119 { 1120 int qentry; 1121 1122 qentry = sq->tail; 1123 if (!sq->is_xdp) 1124 atomic_sub(desc_cnt, &sq->free_cnt); 1125 else 1126 sq->xdp_free_cnt -= desc_cnt; 1127 sq->tail += desc_cnt; 1128 sq->tail &= (sq->dmem.q_len - 1); 1129 1130 return qentry; 1131 } 1132 1133 /* Rollback to previous tail pointer when descriptors not used */ 1134 static inline void nicvf_rollback_sq_desc(struct snd_queue *sq, 1135 int qentry, int desc_cnt) 1136 { 1137 sq->tail = qentry; 1138 atomic_add(desc_cnt, &sq->free_cnt); 1139 } 1140 1141 /* Free descriptor back to SQ for future use */ 1142 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt) 1143 { 1144 if (!sq->is_xdp) 1145 atomic_add(desc_cnt, &sq->free_cnt); 1146 else 1147 sq->xdp_free_cnt += desc_cnt; 1148 sq->head += desc_cnt; 1149 sq->head &= (sq->dmem.q_len - 1); 1150 } 1151 1152 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry) 1153 { 1154 qentry++; 1155 qentry &= (sq->dmem.q_len - 1); 1156 return qentry; 1157 } 1158 1159 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx) 1160 { 1161 u64 sq_cfg; 1162 1163 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1164 sq_cfg |= NICVF_SQ_EN; 1165 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1166 /* Ring doorbell so that H/W restarts processing SQEs */ 1167 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0); 1168 } 1169 1170 void nicvf_sq_disable(struct nicvf *nic, int qidx) 1171 { 1172 u64 sq_cfg; 1173 1174 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx); 1175 sq_cfg &= ~NICVF_SQ_EN; 1176 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg); 1177 } 1178 1179 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq, 1180 int qidx) 1181 { 1182 u64 head, tail; 1183 struct sk_buff *skb; 1184 struct nicvf *nic = netdev_priv(netdev); 1185 struct sq_hdr_subdesc *hdr; 1186 1187 head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4; 1188 tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4; 1189 while (sq->head != head) { 1190 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head); 1191 if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) { 1192 nicvf_put_sq_desc(sq, 1); 1193 continue; 1194 } 1195 skb = (struct sk_buff *)sq->skbuff[sq->head]; 1196 if (skb) 1197 dev_kfree_skb_any(skb); 1198 atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets); 1199 atomic64_add(hdr->tot_len, 1200 (atomic64_t *)&netdev->stats.tx_bytes); 1201 nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1); 1202 } 1203 } 1204 1205 /* XDP Transmit APIs */ 1206 void nicvf_xdp_sq_doorbell(struct nicvf *nic, 1207 struct snd_queue *sq, int sq_num) 1208 { 1209 if (!sq->xdp_desc_cnt) 1210 return; 1211 1212 /* make sure all memory stores are done before ringing doorbell */ 1213 wmb(); 1214 1215 /* Inform HW to xmit all TSO segments */ 1216 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, 1217 sq_num, sq->xdp_desc_cnt); 1218 sq->xdp_desc_cnt = 0; 1219 } 1220 1221 static inline void 1222 nicvf_xdp_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry, 1223 int subdesc_cnt, u64 data, int len) 1224 { 1225 struct sq_hdr_subdesc *hdr; 1226 1227 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry); 1228 memset(hdr, 0, SND_QUEUE_DESC_SIZE); 1229 hdr->subdesc_type = SQ_DESC_TYPE_HEADER; 1230 hdr->subdesc_cnt = subdesc_cnt; 1231 hdr->tot_len = len; 1232 hdr->post_cqe = 1; 1233 sq->xdp_page[qentry] = (u64)virt_to_page((void *)data); 1234 } 1235 1236 int nicvf_xdp_sq_append_pkt(struct nicvf *nic, struct snd_queue *sq, 1237 u64 bufaddr, u64 dma_addr, u16 len) 1238 { 1239 int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT; 1240 int qentry; 1241 1242 if (subdesc_cnt > sq->xdp_free_cnt) 1243 return 0; 1244 1245 qentry = nicvf_get_sq_desc(sq, subdesc_cnt); 1246 1247 nicvf_xdp_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, bufaddr, len); 1248 1249 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1250 nicvf_sq_add_gather_subdesc(sq, qentry, len, dma_addr); 1251 1252 sq->xdp_desc_cnt += subdesc_cnt; 1253 1254 return 1; 1255 } 1256 1257 /* Calculate no of SQ subdescriptors needed to transmit all 1258 * segments of this TSO packet. 1259 * Taken from 'Tilera network driver' with a minor modification. 1260 */ 1261 static int nicvf_tso_count_subdescs(struct sk_buff *skb) 1262 { 1263 struct skb_shared_info *sh = skb_shinfo(skb); 1264 unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb); 1265 unsigned int data_len = skb->len - sh_len; 1266 unsigned int p_len = sh->gso_size; 1267 long f_id = -1; /* id of the current fragment */ 1268 long f_size = skb_headlen(skb) - sh_len; /* current fragment size */ 1269 long f_used = 0; /* bytes used from the current fragment */ 1270 long n; /* size of the current piece of payload */ 1271 int num_edescs = 0; 1272 int segment; 1273 1274 for (segment = 0; segment < sh->gso_segs; segment++) { 1275 unsigned int p_used = 0; 1276 1277 /* One edesc for header and for each piece of the payload. */ 1278 for (num_edescs++; p_used < p_len; num_edescs++) { 1279 /* Advance as needed. */ 1280 while (f_used >= f_size) { 1281 f_id++; 1282 f_size = skb_frag_size(&sh->frags[f_id]); 1283 f_used = 0; 1284 } 1285 1286 /* Use bytes from the current fragment. */ 1287 n = p_len - p_used; 1288 if (n > f_size - f_used) 1289 n = f_size - f_used; 1290 f_used += n; 1291 p_used += n; 1292 } 1293 1294 /* The last segment may be less than gso_size. */ 1295 data_len -= p_len; 1296 if (data_len < p_len) 1297 p_len = data_len; 1298 } 1299 1300 /* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */ 1301 return num_edescs + sh->gso_segs; 1302 } 1303 1304 #define POST_CQE_DESC_COUNT 2 1305 1306 /* Get the number of SQ descriptors needed to xmit this skb */ 1307 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb) 1308 { 1309 int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT; 1310 1311 if (skb_shinfo(skb)->gso_size && !nic->hw_tso) { 1312 subdesc_cnt = nicvf_tso_count_subdescs(skb); 1313 return subdesc_cnt; 1314 } 1315 1316 /* Dummy descriptors to get TSO pkt completion notification */ 1317 if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) 1318 subdesc_cnt += POST_CQE_DESC_COUNT; 1319 1320 if (skb_shinfo(skb)->nr_frags) 1321 subdesc_cnt += skb_shinfo(skb)->nr_frags; 1322 1323 return subdesc_cnt; 1324 } 1325 1326 /* Add SQ HEADER subdescriptor. 1327 * First subdescriptor for every send descriptor. 1328 */ 1329 static inline void 1330 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry, 1331 int subdesc_cnt, struct sk_buff *skb, int len) 1332 { 1333 int proto; 1334 struct sq_hdr_subdesc *hdr; 1335 union { 1336 struct iphdr *v4; 1337 struct ipv6hdr *v6; 1338 unsigned char *hdr; 1339 } ip; 1340 1341 ip.hdr = skb_network_header(skb); 1342 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry); 1343 memset(hdr, 0, SND_QUEUE_DESC_SIZE); 1344 hdr->subdesc_type = SQ_DESC_TYPE_HEADER; 1345 1346 if (nic->t88 && nic->hw_tso && skb_shinfo(skb)->gso_size) { 1347 /* post_cqe = 0, to avoid HW posting a CQE for every TSO 1348 * segment transmitted on 88xx. 1349 */ 1350 hdr->subdesc_cnt = subdesc_cnt - POST_CQE_DESC_COUNT; 1351 } else { 1352 sq->skbuff[qentry] = (u64)skb; 1353 /* Enable notification via CQE after processing SQE */ 1354 hdr->post_cqe = 1; 1355 /* No of subdescriptors following this */ 1356 hdr->subdesc_cnt = subdesc_cnt; 1357 } 1358 hdr->tot_len = len; 1359 1360 /* Offload checksum calculation to HW */ 1361 if (skb->ip_summed == CHECKSUM_PARTIAL) { 1362 if (ip.v4->version == 4) 1363 hdr->csum_l3 = 1; /* Enable IP csum calculation */ 1364 hdr->l3_offset = skb_network_offset(skb); 1365 hdr->l4_offset = skb_transport_offset(skb); 1366 1367 proto = (ip.v4->version == 4) ? ip.v4->protocol : 1368 ip.v6->nexthdr; 1369 1370 switch (proto) { 1371 case IPPROTO_TCP: 1372 hdr->csum_l4 = SEND_L4_CSUM_TCP; 1373 break; 1374 case IPPROTO_UDP: 1375 hdr->csum_l4 = SEND_L4_CSUM_UDP; 1376 break; 1377 case IPPROTO_SCTP: 1378 hdr->csum_l4 = SEND_L4_CSUM_SCTP; 1379 break; 1380 } 1381 } 1382 1383 if (nic->hw_tso && skb_shinfo(skb)->gso_size) { 1384 hdr->tso = 1; 1385 hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb); 1386 hdr->tso_max_paysize = skb_shinfo(skb)->gso_size; 1387 /* For non-tunneled pkts, point this to L2 ethertype */ 1388 hdr->inner_l3_offset = skb_network_offset(skb) - 2; 1389 this_cpu_inc(nic->pnicvf->drv_stats->tx_tso); 1390 } 1391 1392 /* Check if timestamp is requested */ 1393 if (!(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)) { 1394 skb_tx_timestamp(skb); 1395 return; 1396 } 1397 1398 /* Tx timestamping not supported along with TSO, so ignore request */ 1399 if (skb_shinfo(skb)->gso_size) 1400 return; 1401 1402 /* HW supports only a single outstanding packet to timestamp */ 1403 if (!atomic_add_unless(&nic->pnicvf->tx_ptp_skbs, 1, 1)) 1404 return; 1405 1406 /* Mark the SKB for later reference */ 1407 skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS; 1408 1409 /* Finally enable timestamp generation 1410 * Since 'post_cqe' is also set, two CQEs will be posted 1411 * for this packet i.e CQE_TYPE_SEND and CQE_TYPE_SEND_PTP. 1412 */ 1413 hdr->tstmp = 1; 1414 } 1415 1416 /* SQ GATHER subdescriptor 1417 * Must follow HDR descriptor 1418 */ 1419 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry, 1420 int size, u64 data) 1421 { 1422 struct sq_gather_subdesc *gather; 1423 1424 qentry &= (sq->dmem.q_len - 1); 1425 gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry); 1426 1427 memset(gather, 0, SND_QUEUE_DESC_SIZE); 1428 gather->subdesc_type = SQ_DESC_TYPE_GATHER; 1429 gather->ld_type = NIC_SEND_LD_TYPE_E_LDD; 1430 gather->size = size; 1431 gather->addr = data; 1432 } 1433 1434 /* Add HDR + IMMEDIATE subdescriptors right after descriptors of a TSO 1435 * packet so that a CQE is posted as a notifation for transmission of 1436 * TSO packet. 1437 */ 1438 static inline void nicvf_sq_add_cqe_subdesc(struct snd_queue *sq, int qentry, 1439 int tso_sqe, struct sk_buff *skb) 1440 { 1441 struct sq_imm_subdesc *imm; 1442 struct sq_hdr_subdesc *hdr; 1443 1444 sq->skbuff[qentry] = (u64)skb; 1445 1446 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry); 1447 memset(hdr, 0, SND_QUEUE_DESC_SIZE); 1448 hdr->subdesc_type = SQ_DESC_TYPE_HEADER; 1449 /* Enable notification via CQE after processing SQE */ 1450 hdr->post_cqe = 1; 1451 /* There is no packet to transmit here */ 1452 hdr->dont_send = 1; 1453 hdr->subdesc_cnt = POST_CQE_DESC_COUNT - 1; 1454 hdr->tot_len = 1; 1455 /* Actual TSO header SQE index, needed for cleanup */ 1456 hdr->rsvd2 = tso_sqe; 1457 1458 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1459 imm = (struct sq_imm_subdesc *)GET_SQ_DESC(sq, qentry); 1460 memset(imm, 0, SND_QUEUE_DESC_SIZE); 1461 imm->subdesc_type = SQ_DESC_TYPE_IMMEDIATE; 1462 imm->len = 1; 1463 } 1464 1465 static inline void nicvf_sq_doorbell(struct nicvf *nic, struct sk_buff *skb, 1466 int sq_num, int desc_cnt) 1467 { 1468 struct netdev_queue *txq; 1469 1470 txq = netdev_get_tx_queue(nic->pnicvf->netdev, 1471 skb_get_queue_mapping(skb)); 1472 1473 netdev_tx_sent_queue(txq, skb->len); 1474 1475 /* make sure all memory stores are done before ringing doorbell */ 1476 smp_wmb(); 1477 1478 /* Inform HW to xmit all TSO segments */ 1479 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, 1480 sq_num, desc_cnt); 1481 } 1482 1483 /* Segment a TSO packet into 'gso_size' segments and append 1484 * them to SQ for transfer 1485 */ 1486 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq, 1487 int sq_num, int qentry, struct sk_buff *skb) 1488 { 1489 struct tso_t tso; 1490 int seg_subdescs = 0, desc_cnt = 0; 1491 int seg_len, total_len, data_left; 1492 int hdr_qentry = qentry; 1493 int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb); 1494 1495 tso_start(skb, &tso); 1496 total_len = skb->len - hdr_len; 1497 while (total_len > 0) { 1498 char *hdr; 1499 1500 /* Save Qentry for adding HDR_SUBDESC at the end */ 1501 hdr_qentry = qentry; 1502 1503 data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len); 1504 total_len -= data_left; 1505 1506 /* Add segment's header */ 1507 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1508 hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE; 1509 tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0); 1510 nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len, 1511 sq->tso_hdrs_phys + 1512 qentry * TSO_HEADER_SIZE); 1513 /* HDR_SUDESC + GATHER */ 1514 seg_subdescs = 2; 1515 seg_len = hdr_len; 1516 1517 /* Add segment's payload fragments */ 1518 while (data_left > 0) { 1519 int size; 1520 1521 size = min_t(int, tso.size, data_left); 1522 1523 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1524 nicvf_sq_add_gather_subdesc(sq, qentry, size, 1525 virt_to_phys(tso.data)); 1526 seg_subdescs++; 1527 seg_len += size; 1528 1529 data_left -= size; 1530 tso_build_data(skb, &tso, size); 1531 } 1532 nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry, 1533 seg_subdescs - 1, skb, seg_len); 1534 sq->skbuff[hdr_qentry] = (u64)NULL; 1535 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1536 1537 desc_cnt += seg_subdescs; 1538 } 1539 /* Save SKB in the last segment for freeing */ 1540 sq->skbuff[hdr_qentry] = (u64)skb; 1541 1542 nicvf_sq_doorbell(nic, skb, sq_num, desc_cnt); 1543 1544 this_cpu_inc(nic->pnicvf->drv_stats->tx_tso); 1545 return 1; 1546 } 1547 1548 /* Append an skb to a SQ for packet transfer. */ 1549 int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq, 1550 struct sk_buff *skb, u8 sq_num) 1551 { 1552 int i, size; 1553 int subdesc_cnt, hdr_sqe = 0; 1554 int qentry; 1555 u64 dma_addr; 1556 1557 subdesc_cnt = nicvf_sq_subdesc_required(nic, skb); 1558 if (subdesc_cnt > atomic_read(&sq->free_cnt)) 1559 goto append_fail; 1560 1561 qentry = nicvf_get_sq_desc(sq, subdesc_cnt); 1562 1563 /* Check if its a TSO packet */ 1564 if (skb_shinfo(skb)->gso_size && !nic->hw_tso) 1565 return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb); 1566 1567 /* Add SQ header subdesc */ 1568 nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1, 1569 skb, skb->len); 1570 hdr_sqe = qentry; 1571 1572 /* Add SQ gather subdescs */ 1573 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1574 size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len; 1575 /* HW will ensure data coherency, CPU sync not required */ 1576 dma_addr = dma_map_page_attrs(&nic->pdev->dev, virt_to_page(skb->data), 1577 offset_in_page(skb->data), size, 1578 DMA_TO_DEVICE, DMA_ATTR_SKIP_CPU_SYNC); 1579 if (dma_mapping_error(&nic->pdev->dev, dma_addr)) { 1580 nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt); 1581 return 0; 1582 } 1583 1584 nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr); 1585 1586 /* Check for scattered buffer */ 1587 if (!skb_is_nonlinear(skb)) 1588 goto doorbell; 1589 1590 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 1591 const struct skb_frag_struct *frag; 1592 1593 frag = &skb_shinfo(skb)->frags[i]; 1594 1595 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1596 size = skb_frag_size(frag); 1597 dma_addr = dma_map_page_attrs(&nic->pdev->dev, 1598 skb_frag_page(frag), 1599 frag->page_offset, size, 1600 DMA_TO_DEVICE, 1601 DMA_ATTR_SKIP_CPU_SYNC); 1602 if (dma_mapping_error(&nic->pdev->dev, dma_addr)) { 1603 /* Free entire chain of mapped buffers 1604 * here 'i' = frags mapped + above mapped skb->data 1605 */ 1606 nicvf_unmap_sndq_buffers(nic, sq, hdr_sqe, i); 1607 nicvf_rollback_sq_desc(sq, qentry, subdesc_cnt); 1608 return 0; 1609 } 1610 nicvf_sq_add_gather_subdesc(sq, qentry, size, dma_addr); 1611 } 1612 1613 doorbell: 1614 if (nic->t88 && skb_shinfo(skb)->gso_size) { 1615 qentry = nicvf_get_nxt_sqentry(sq, qentry); 1616 nicvf_sq_add_cqe_subdesc(sq, qentry, hdr_sqe, skb); 1617 } 1618 1619 nicvf_sq_doorbell(nic, skb, sq_num, subdesc_cnt); 1620 1621 return 1; 1622 1623 append_fail: 1624 /* Use original PCI dev for debug log */ 1625 nic = nic->pnicvf; 1626 netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n"); 1627 return 0; 1628 } 1629 1630 static inline unsigned frag_num(unsigned i) 1631 { 1632 #ifdef __BIG_ENDIAN 1633 return (i & ~3) + 3 - (i & 3); 1634 #else 1635 return i; 1636 #endif 1637 } 1638 1639 static void nicvf_unmap_rcv_buffer(struct nicvf *nic, u64 dma_addr, 1640 u64 buf_addr, bool xdp) 1641 { 1642 struct page *page = NULL; 1643 int len = RCV_FRAG_LEN; 1644 1645 if (xdp) { 1646 page = virt_to_page(phys_to_virt(buf_addr)); 1647 /* Check if it's a recycled page, if not 1648 * unmap the DMA mapping. 1649 * 1650 * Recycled page holds an extra reference. 1651 */ 1652 if (page_ref_count(page) != 1) 1653 return; 1654 1655 len += XDP_PACKET_HEADROOM; 1656 /* Receive buffers in XDP mode are mapped from page start */ 1657 dma_addr &= PAGE_MASK; 1658 } 1659 dma_unmap_page_attrs(&nic->pdev->dev, dma_addr, len, 1660 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC); 1661 } 1662 1663 /* Returns SKB for a received packet */ 1664 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic, 1665 struct cqe_rx_t *cqe_rx, bool xdp) 1666 { 1667 int frag; 1668 int payload_len = 0; 1669 struct sk_buff *skb = NULL; 1670 struct page *page; 1671 int offset; 1672 u16 *rb_lens = NULL; 1673 u64 *rb_ptrs = NULL; 1674 u64 phys_addr; 1675 1676 rb_lens = (void *)cqe_rx + (3 * sizeof(u64)); 1677 /* Except 88xx pass1 on all other chips CQE_RX2_S is added to 1678 * CQE_RX at word6, hence buffer pointers move by word 1679 * 1680 * Use existing 'hw_tso' flag which will be set for all chips 1681 * except 88xx pass1 instead of a additional cache line 1682 * access (or miss) by using pci dev's revision. 1683 */ 1684 if (!nic->hw_tso) 1685 rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64)); 1686 else 1687 rb_ptrs = (void *)cqe_rx + (7 * sizeof(u64)); 1688 1689 for (frag = 0; frag < cqe_rx->rb_cnt; frag++) { 1690 payload_len = rb_lens[frag_num(frag)]; 1691 phys_addr = nicvf_iova_to_phys(nic, *rb_ptrs); 1692 if (!phys_addr) { 1693 if (skb) 1694 dev_kfree_skb_any(skb); 1695 return NULL; 1696 } 1697 1698 if (!frag) { 1699 /* First fragment */ 1700 nicvf_unmap_rcv_buffer(nic, 1701 *rb_ptrs - cqe_rx->align_pad, 1702 phys_addr, xdp); 1703 skb = nicvf_rb_ptr_to_skb(nic, 1704 phys_addr - cqe_rx->align_pad, 1705 payload_len); 1706 if (!skb) 1707 return NULL; 1708 skb_reserve(skb, cqe_rx->align_pad); 1709 skb_put(skb, payload_len); 1710 } else { 1711 /* Add fragments */ 1712 nicvf_unmap_rcv_buffer(nic, *rb_ptrs, phys_addr, xdp); 1713 page = virt_to_page(phys_to_virt(phys_addr)); 1714 offset = phys_to_virt(phys_addr) - page_address(page); 1715 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, 1716 offset, payload_len, RCV_FRAG_LEN); 1717 } 1718 /* Next buffer pointer */ 1719 rb_ptrs++; 1720 } 1721 return skb; 1722 } 1723 1724 static u64 nicvf_int_type_to_mask(int int_type, int q_idx) 1725 { 1726 u64 reg_val; 1727 1728 switch (int_type) { 1729 case NICVF_INTR_CQ: 1730 reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT); 1731 break; 1732 case NICVF_INTR_SQ: 1733 reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT); 1734 break; 1735 case NICVF_INTR_RBDR: 1736 reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT); 1737 break; 1738 case NICVF_INTR_PKT_DROP: 1739 reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT); 1740 break; 1741 case NICVF_INTR_TCP_TIMER: 1742 reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT); 1743 break; 1744 case NICVF_INTR_MBOX: 1745 reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT); 1746 break; 1747 case NICVF_INTR_QS_ERR: 1748 reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT); 1749 break; 1750 default: 1751 reg_val = 0; 1752 } 1753 1754 return reg_val; 1755 } 1756 1757 /* Enable interrupt */ 1758 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx) 1759 { 1760 u64 mask = nicvf_int_type_to_mask(int_type, q_idx); 1761 1762 if (!mask) { 1763 netdev_dbg(nic->netdev, 1764 "Failed to enable interrupt: unknown type\n"); 1765 return; 1766 } 1767 nicvf_reg_write(nic, NIC_VF_ENA_W1S, 1768 nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask); 1769 } 1770 1771 /* Disable interrupt */ 1772 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx) 1773 { 1774 u64 mask = nicvf_int_type_to_mask(int_type, q_idx); 1775 1776 if (!mask) { 1777 netdev_dbg(nic->netdev, 1778 "Failed to disable interrupt: unknown type\n"); 1779 return; 1780 } 1781 1782 nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask); 1783 } 1784 1785 /* Clear interrupt */ 1786 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx) 1787 { 1788 u64 mask = nicvf_int_type_to_mask(int_type, q_idx); 1789 1790 if (!mask) { 1791 netdev_dbg(nic->netdev, 1792 "Failed to clear interrupt: unknown type\n"); 1793 return; 1794 } 1795 1796 nicvf_reg_write(nic, NIC_VF_INT, mask); 1797 } 1798 1799 /* Check if interrupt is enabled */ 1800 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx) 1801 { 1802 u64 mask = nicvf_int_type_to_mask(int_type, q_idx); 1803 /* If interrupt type is unknown, we treat it disabled. */ 1804 if (!mask) { 1805 netdev_dbg(nic->netdev, 1806 "Failed to check interrupt enable: unknown type\n"); 1807 return 0; 1808 } 1809 1810 return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S); 1811 } 1812 1813 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx) 1814 { 1815 struct rcv_queue *rq; 1816 1817 #define GET_RQ_STATS(reg) \ 1818 nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\ 1819 (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 1820 1821 rq = &nic->qs->rq[rq_idx]; 1822 rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS); 1823 rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS); 1824 } 1825 1826 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx) 1827 { 1828 struct snd_queue *sq; 1829 1830 #define GET_SQ_STATS(reg) \ 1831 nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\ 1832 (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3)) 1833 1834 sq = &nic->qs->sq[sq_idx]; 1835 sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS); 1836 sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS); 1837 } 1838 1839 /* Check for errors in the receive cmp.queue entry */ 1840 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx) 1841 { 1842 netif_err(nic, rx_err, nic->netdev, 1843 "RX error CQE err_level 0x%x err_opcode 0x%x\n", 1844 cqe_rx->err_level, cqe_rx->err_opcode); 1845 1846 switch (cqe_rx->err_opcode) { 1847 case CQ_RX_ERROP_RE_PARTIAL: 1848 this_cpu_inc(nic->drv_stats->rx_bgx_truncated_pkts); 1849 break; 1850 case CQ_RX_ERROP_RE_JABBER: 1851 this_cpu_inc(nic->drv_stats->rx_jabber_errs); 1852 break; 1853 case CQ_RX_ERROP_RE_FCS: 1854 this_cpu_inc(nic->drv_stats->rx_fcs_errs); 1855 break; 1856 case CQ_RX_ERROP_RE_RX_CTL: 1857 this_cpu_inc(nic->drv_stats->rx_bgx_errs); 1858 break; 1859 case CQ_RX_ERROP_PREL2_ERR: 1860 this_cpu_inc(nic->drv_stats->rx_prel2_errs); 1861 break; 1862 case CQ_RX_ERROP_L2_MAL: 1863 this_cpu_inc(nic->drv_stats->rx_l2_hdr_malformed); 1864 break; 1865 case CQ_RX_ERROP_L2_OVERSIZE: 1866 this_cpu_inc(nic->drv_stats->rx_oversize); 1867 break; 1868 case CQ_RX_ERROP_L2_UNDERSIZE: 1869 this_cpu_inc(nic->drv_stats->rx_undersize); 1870 break; 1871 case CQ_RX_ERROP_L2_LENMISM: 1872 this_cpu_inc(nic->drv_stats->rx_l2_len_mismatch); 1873 break; 1874 case CQ_RX_ERROP_L2_PCLP: 1875 this_cpu_inc(nic->drv_stats->rx_l2_pclp); 1876 break; 1877 case CQ_RX_ERROP_IP_NOT: 1878 this_cpu_inc(nic->drv_stats->rx_ip_ver_errs); 1879 break; 1880 case CQ_RX_ERROP_IP_CSUM_ERR: 1881 this_cpu_inc(nic->drv_stats->rx_ip_csum_errs); 1882 break; 1883 case CQ_RX_ERROP_IP_MAL: 1884 this_cpu_inc(nic->drv_stats->rx_ip_hdr_malformed); 1885 break; 1886 case CQ_RX_ERROP_IP_MALD: 1887 this_cpu_inc(nic->drv_stats->rx_ip_payload_malformed); 1888 break; 1889 case CQ_RX_ERROP_IP_HOP: 1890 this_cpu_inc(nic->drv_stats->rx_ip_ttl_errs); 1891 break; 1892 case CQ_RX_ERROP_L3_PCLP: 1893 this_cpu_inc(nic->drv_stats->rx_l3_pclp); 1894 break; 1895 case CQ_RX_ERROP_L4_MAL: 1896 this_cpu_inc(nic->drv_stats->rx_l4_malformed); 1897 break; 1898 case CQ_RX_ERROP_L4_CHK: 1899 this_cpu_inc(nic->drv_stats->rx_l4_csum_errs); 1900 break; 1901 case CQ_RX_ERROP_UDP_LEN: 1902 this_cpu_inc(nic->drv_stats->rx_udp_len_errs); 1903 break; 1904 case CQ_RX_ERROP_L4_PORT: 1905 this_cpu_inc(nic->drv_stats->rx_l4_port_errs); 1906 break; 1907 case CQ_RX_ERROP_TCP_FLAG: 1908 this_cpu_inc(nic->drv_stats->rx_tcp_flag_errs); 1909 break; 1910 case CQ_RX_ERROP_TCP_OFFSET: 1911 this_cpu_inc(nic->drv_stats->rx_tcp_offset_errs); 1912 break; 1913 case CQ_RX_ERROP_L4_PCLP: 1914 this_cpu_inc(nic->drv_stats->rx_l4_pclp); 1915 break; 1916 case CQ_RX_ERROP_RBDR_TRUNC: 1917 this_cpu_inc(nic->drv_stats->rx_truncated_pkts); 1918 break; 1919 } 1920 1921 return 1; 1922 } 1923 1924 /* Check for errors in the send cmp.queue entry */ 1925 int nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cqe_send_t *cqe_tx) 1926 { 1927 switch (cqe_tx->send_status) { 1928 case CQ_TX_ERROP_DESC_FAULT: 1929 this_cpu_inc(nic->drv_stats->tx_desc_fault); 1930 break; 1931 case CQ_TX_ERROP_HDR_CONS_ERR: 1932 this_cpu_inc(nic->drv_stats->tx_hdr_cons_err); 1933 break; 1934 case CQ_TX_ERROP_SUBDC_ERR: 1935 this_cpu_inc(nic->drv_stats->tx_subdesc_err); 1936 break; 1937 case CQ_TX_ERROP_MAX_SIZE_VIOL: 1938 this_cpu_inc(nic->drv_stats->tx_max_size_exceeded); 1939 break; 1940 case CQ_TX_ERROP_IMM_SIZE_OFLOW: 1941 this_cpu_inc(nic->drv_stats->tx_imm_size_oflow); 1942 break; 1943 case CQ_TX_ERROP_DATA_SEQUENCE_ERR: 1944 this_cpu_inc(nic->drv_stats->tx_data_seq_err); 1945 break; 1946 case CQ_TX_ERROP_MEM_SEQUENCE_ERR: 1947 this_cpu_inc(nic->drv_stats->tx_mem_seq_err); 1948 break; 1949 case CQ_TX_ERROP_LOCK_VIOL: 1950 this_cpu_inc(nic->drv_stats->tx_lock_viol); 1951 break; 1952 case CQ_TX_ERROP_DATA_FAULT: 1953 this_cpu_inc(nic->drv_stats->tx_data_fault); 1954 break; 1955 case CQ_TX_ERROP_TSTMP_CONFLICT: 1956 this_cpu_inc(nic->drv_stats->tx_tstmp_conflict); 1957 break; 1958 case CQ_TX_ERROP_TSTMP_TIMEOUT: 1959 this_cpu_inc(nic->drv_stats->tx_tstmp_timeout); 1960 break; 1961 case CQ_TX_ERROP_MEM_FAULT: 1962 this_cpu_inc(nic->drv_stats->tx_mem_fault); 1963 break; 1964 case CQ_TX_ERROP_CK_OVERLAP: 1965 this_cpu_inc(nic->drv_stats->tx_csum_overlap); 1966 break; 1967 case CQ_TX_ERROP_CK_OFLOW: 1968 this_cpu_inc(nic->drv_stats->tx_csum_overflow); 1969 break; 1970 } 1971 1972 return 1; 1973 } 1974