1 // SPDX-License-Identifier: ISC 2 /* 3 * Copyright (C) 2016 Felix Fietkau <nbd@nbd.name> 4 */ 5 6 #include <linux/dma-mapping.h> 7 #include "mt76.h" 8 #include "dma.h" 9 10 static struct mt76_txwi_cache * 11 mt76_alloc_txwi(struct mt76_dev *dev) 12 { 13 struct mt76_txwi_cache *t; 14 dma_addr_t addr; 15 u8 *txwi; 16 int size; 17 18 size = L1_CACHE_ALIGN(dev->drv->txwi_size + sizeof(*t)); 19 txwi = devm_kzalloc(dev->dev, size, GFP_ATOMIC); 20 if (!txwi) 21 return NULL; 22 23 addr = dma_map_single(dev->dev, txwi, dev->drv->txwi_size, 24 DMA_TO_DEVICE); 25 t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size); 26 t->dma_addr = addr; 27 28 return t; 29 } 30 31 static struct mt76_txwi_cache * 32 __mt76_get_txwi(struct mt76_dev *dev) 33 { 34 struct mt76_txwi_cache *t = NULL; 35 36 spin_lock(&dev->lock); 37 if (!list_empty(&dev->txwi_cache)) { 38 t = list_first_entry(&dev->txwi_cache, struct mt76_txwi_cache, 39 list); 40 list_del(&t->list); 41 } 42 spin_unlock(&dev->lock); 43 44 return t; 45 } 46 47 static struct mt76_txwi_cache * 48 mt76_get_txwi(struct mt76_dev *dev) 49 { 50 struct mt76_txwi_cache *t = __mt76_get_txwi(dev); 51 52 if (t) 53 return t; 54 55 return mt76_alloc_txwi(dev); 56 } 57 58 void 59 mt76_put_txwi(struct mt76_dev *dev, struct mt76_txwi_cache *t) 60 { 61 if (!t) 62 return; 63 64 spin_lock(&dev->lock); 65 list_add(&t->list, &dev->txwi_cache); 66 spin_unlock(&dev->lock); 67 } 68 EXPORT_SYMBOL_GPL(mt76_put_txwi); 69 70 static void 71 mt76_free_pending_txwi(struct mt76_dev *dev) 72 { 73 struct mt76_txwi_cache *t; 74 75 local_bh_disable(); 76 while ((t = __mt76_get_txwi(dev)) != NULL) 77 dma_unmap_single(dev->dev, t->dma_addr, dev->drv->txwi_size, 78 DMA_TO_DEVICE); 79 local_bh_enable(); 80 } 81 82 static int 83 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q, 84 int idx, int n_desc, int bufsize, 85 u32 ring_base) 86 { 87 int size; 88 int i; 89 90 spin_lock_init(&q->lock); 91 spin_lock_init(&q->cleanup_lock); 92 93 q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE; 94 q->ndesc = n_desc; 95 q->buf_size = bufsize; 96 q->hw_idx = idx; 97 98 size = q->ndesc * sizeof(struct mt76_desc); 99 q->desc = dmam_alloc_coherent(dev->dev, size, &q->desc_dma, GFP_KERNEL); 100 if (!q->desc) 101 return -ENOMEM; 102 103 size = q->ndesc * sizeof(*q->entry); 104 q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL); 105 if (!q->entry) 106 return -ENOMEM; 107 108 /* clear descriptors */ 109 for (i = 0; i < q->ndesc; i++) 110 q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE); 111 112 writel(q->desc_dma, &q->regs->desc_base); 113 writel(0, &q->regs->cpu_idx); 114 writel(0, &q->regs->dma_idx); 115 writel(q->ndesc, &q->regs->ring_size); 116 117 return 0; 118 } 119 120 static int 121 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q, 122 struct mt76_queue_buf *buf, int nbufs, u32 info, 123 struct sk_buff *skb, void *txwi) 124 { 125 struct mt76_queue_entry *entry; 126 struct mt76_desc *desc; 127 u32 ctrl; 128 int i, idx = -1; 129 130 if (txwi) { 131 q->entry[q->head].txwi = DMA_DUMMY_DATA; 132 q->entry[q->head].skip_buf0 = true; 133 } 134 135 for (i = 0; i < nbufs; i += 2, buf += 2) { 136 u32 buf0 = buf[0].addr, buf1 = 0; 137 138 idx = q->head; 139 q->head = (q->head + 1) % q->ndesc; 140 141 desc = &q->desc[idx]; 142 entry = &q->entry[idx]; 143 144 if (buf[0].skip_unmap) 145 entry->skip_buf0 = true; 146 entry->skip_buf1 = i == nbufs - 1; 147 148 entry->dma_addr[0] = buf[0].addr; 149 entry->dma_len[0] = buf[0].len; 150 151 ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len); 152 if (i < nbufs - 1) { 153 entry->dma_addr[1] = buf[1].addr; 154 entry->dma_len[1] = buf[1].len; 155 buf1 = buf[1].addr; 156 ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len); 157 if (buf[1].skip_unmap) 158 entry->skip_buf1 = true; 159 } 160 161 if (i == nbufs - 1) 162 ctrl |= MT_DMA_CTL_LAST_SEC0; 163 else if (i == nbufs - 2) 164 ctrl |= MT_DMA_CTL_LAST_SEC1; 165 166 WRITE_ONCE(desc->buf0, cpu_to_le32(buf0)); 167 WRITE_ONCE(desc->buf1, cpu_to_le32(buf1)); 168 WRITE_ONCE(desc->info, cpu_to_le32(info)); 169 WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl)); 170 171 q->queued++; 172 } 173 174 q->entry[idx].txwi = txwi; 175 q->entry[idx].skb = skb; 176 177 return idx; 178 } 179 180 static void 181 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx, 182 struct mt76_queue_entry *prev_e) 183 { 184 struct mt76_queue_entry *e = &q->entry[idx]; 185 186 if (!e->skip_buf0) 187 dma_unmap_single(dev->dev, e->dma_addr[0], e->dma_len[0], 188 DMA_TO_DEVICE); 189 190 if (!e->skip_buf1) 191 dma_unmap_single(dev->dev, e->dma_addr[1], e->dma_len[1], 192 DMA_TO_DEVICE); 193 194 if (e->txwi == DMA_DUMMY_DATA) 195 e->txwi = NULL; 196 197 if (e->skb == DMA_DUMMY_DATA) 198 e->skb = NULL; 199 200 *prev_e = *e; 201 memset(e, 0, sizeof(*e)); 202 } 203 204 static void 205 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q) 206 { 207 writel(q->desc_dma, &q->regs->desc_base); 208 writel(q->ndesc, &q->regs->ring_size); 209 q->head = readl(&q->regs->dma_idx); 210 q->tail = q->head; 211 } 212 213 static void 214 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q) 215 { 216 wmb(); 217 writel(q->head, &q->regs->cpu_idx); 218 } 219 220 static void 221 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush) 222 { 223 struct mt76_queue_entry entry; 224 int last; 225 226 if (!q) 227 return; 228 229 spin_lock_bh(&q->cleanup_lock); 230 if (flush) 231 last = -1; 232 else 233 last = readl(&q->regs->dma_idx); 234 235 while (q->queued > 0 && q->tail != last) { 236 mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry); 237 mt76_queue_tx_complete(dev, q, &entry); 238 239 if (entry.txwi) { 240 if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE)) 241 mt76_put_txwi(dev, entry.txwi); 242 } 243 244 if (!flush && q->tail == last) 245 last = readl(&q->regs->dma_idx); 246 247 } 248 spin_unlock_bh(&q->cleanup_lock); 249 250 if (flush) { 251 spin_lock_bh(&q->lock); 252 mt76_dma_sync_idx(dev, q); 253 mt76_dma_kick_queue(dev, q); 254 spin_unlock_bh(&q->lock); 255 } 256 257 if (!q->queued) 258 wake_up(&dev->tx_wait); 259 } 260 261 static void * 262 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx, 263 int *len, u32 *info, bool *more) 264 { 265 struct mt76_queue_entry *e = &q->entry[idx]; 266 struct mt76_desc *desc = &q->desc[idx]; 267 dma_addr_t buf_addr; 268 void *buf = e->buf; 269 int buf_len = SKB_WITH_OVERHEAD(q->buf_size); 270 271 buf_addr = e->dma_addr[0]; 272 if (len) { 273 u32 ctl = le32_to_cpu(READ_ONCE(desc->ctrl)); 274 *len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctl); 275 *more = !(ctl & MT_DMA_CTL_LAST_SEC0); 276 } 277 278 if (info) 279 *info = le32_to_cpu(desc->info); 280 281 dma_unmap_single(dev->dev, buf_addr, buf_len, DMA_FROM_DEVICE); 282 e->buf = NULL; 283 284 return buf; 285 } 286 287 static void * 288 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush, 289 int *len, u32 *info, bool *more) 290 { 291 int idx = q->tail; 292 293 *more = false; 294 if (!q->queued) 295 return NULL; 296 297 if (flush) 298 q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE); 299 else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE))) 300 return NULL; 301 302 q->tail = (q->tail + 1) % q->ndesc; 303 q->queued--; 304 305 return mt76_dma_get_buf(dev, q, idx, len, info, more); 306 } 307 308 static int 309 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q, 310 struct sk_buff *skb, u32 tx_info) 311 { 312 struct mt76_queue_buf buf; 313 dma_addr_t addr; 314 315 if (q->queued + 1 >= q->ndesc - 1) 316 goto error; 317 318 addr = dma_map_single(dev->dev, skb->data, skb->len, 319 DMA_TO_DEVICE); 320 if (unlikely(dma_mapping_error(dev->dev, addr))) 321 goto error; 322 323 buf.addr = addr; 324 buf.len = skb->len; 325 326 spin_lock_bh(&q->lock); 327 mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL); 328 mt76_dma_kick_queue(dev, q); 329 spin_unlock_bh(&q->lock); 330 331 return 0; 332 333 error: 334 dev_kfree_skb(skb); 335 return -ENOMEM; 336 } 337 338 static int 339 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q, 340 struct sk_buff *skb, struct mt76_wcid *wcid, 341 struct ieee80211_sta *sta) 342 { 343 struct mt76_tx_info tx_info = { 344 .skb = skb, 345 }; 346 struct ieee80211_hw *hw; 347 int len, n = 0, ret = -ENOMEM; 348 struct mt76_txwi_cache *t; 349 struct sk_buff *iter; 350 dma_addr_t addr; 351 u8 *txwi; 352 353 t = mt76_get_txwi(dev); 354 if (!t) { 355 hw = mt76_tx_status_get_hw(dev, skb); 356 ieee80211_free_txskb(hw, skb); 357 return -ENOMEM; 358 } 359 txwi = mt76_get_txwi_ptr(dev, t); 360 361 skb->prev = skb->next = NULL; 362 if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS) 363 mt76_insert_hdr_pad(skb); 364 365 len = skb_headlen(skb); 366 addr = dma_map_single(dev->dev, skb->data, len, DMA_TO_DEVICE); 367 if (unlikely(dma_mapping_error(dev->dev, addr))) 368 goto free; 369 370 tx_info.buf[n].addr = t->dma_addr; 371 tx_info.buf[n++].len = dev->drv->txwi_size; 372 tx_info.buf[n].addr = addr; 373 tx_info.buf[n++].len = len; 374 375 skb_walk_frags(skb, iter) { 376 if (n == ARRAY_SIZE(tx_info.buf)) 377 goto unmap; 378 379 addr = dma_map_single(dev->dev, iter->data, iter->len, 380 DMA_TO_DEVICE); 381 if (unlikely(dma_mapping_error(dev->dev, addr))) 382 goto unmap; 383 384 tx_info.buf[n].addr = addr; 385 tx_info.buf[n++].len = iter->len; 386 } 387 tx_info.nbuf = n; 388 389 if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) { 390 ret = -ENOMEM; 391 goto unmap; 392 } 393 394 dma_sync_single_for_cpu(dev->dev, t->dma_addr, dev->drv->txwi_size, 395 DMA_TO_DEVICE); 396 ret = dev->drv->tx_prepare_skb(dev, txwi, q->qid, wcid, sta, &tx_info); 397 dma_sync_single_for_device(dev->dev, t->dma_addr, dev->drv->txwi_size, 398 DMA_TO_DEVICE); 399 if (ret < 0) 400 goto unmap; 401 402 return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf, 403 tx_info.info, tx_info.skb, t); 404 405 unmap: 406 for (n--; n > 0; n--) 407 dma_unmap_single(dev->dev, tx_info.buf[n].addr, 408 tx_info.buf[n].len, DMA_TO_DEVICE); 409 410 free: 411 #ifdef CONFIG_NL80211_TESTMODE 412 /* fix tx_done accounting on queue overflow */ 413 if (mt76_is_testmode_skb(dev, skb, &hw)) { 414 struct mt76_phy *phy = hw->priv; 415 416 if (tx_info.skb == phy->test.tx_skb) 417 phy->test.tx_done--; 418 } 419 #endif 420 421 dev_kfree_skb(tx_info.skb); 422 mt76_put_txwi(dev, t); 423 return ret; 424 } 425 426 static int 427 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q) 428 { 429 dma_addr_t addr; 430 void *buf; 431 int frames = 0; 432 int len = SKB_WITH_OVERHEAD(q->buf_size); 433 int offset = q->buf_offset; 434 435 spin_lock_bh(&q->lock); 436 437 while (q->queued < q->ndesc - 1) { 438 struct mt76_queue_buf qbuf; 439 440 buf = page_frag_alloc(&q->rx_page, q->buf_size, GFP_ATOMIC); 441 if (!buf) 442 break; 443 444 addr = dma_map_single(dev->dev, buf, len, DMA_FROM_DEVICE); 445 if (unlikely(dma_mapping_error(dev->dev, addr))) { 446 skb_free_frag(buf); 447 break; 448 } 449 450 qbuf.addr = addr + offset; 451 qbuf.len = len - offset; 452 mt76_dma_add_buf(dev, q, &qbuf, 1, 0, buf, NULL); 453 frames++; 454 } 455 456 if (frames) 457 mt76_dma_kick_queue(dev, q); 458 459 spin_unlock_bh(&q->lock); 460 461 return frames; 462 } 463 464 static void 465 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q) 466 { 467 struct page *page; 468 void *buf; 469 bool more; 470 471 spin_lock_bh(&q->lock); 472 do { 473 buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more); 474 if (!buf) 475 break; 476 477 skb_free_frag(buf); 478 } while (1); 479 spin_unlock_bh(&q->lock); 480 481 if (!q->rx_page.va) 482 return; 483 484 page = virt_to_page(q->rx_page.va); 485 __page_frag_cache_drain(page, q->rx_page.pagecnt_bias); 486 memset(&q->rx_page, 0, sizeof(q->rx_page)); 487 } 488 489 static void 490 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid) 491 { 492 struct mt76_queue *q = &dev->q_rx[qid]; 493 int i; 494 495 for (i = 0; i < q->ndesc; i++) 496 q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE); 497 498 mt76_dma_rx_cleanup(dev, q); 499 mt76_dma_sync_idx(dev, q); 500 mt76_dma_rx_fill(dev, q); 501 502 if (!q->rx_head) 503 return; 504 505 dev_kfree_skb(q->rx_head); 506 q->rx_head = NULL; 507 } 508 509 static void 510 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data, 511 int len, bool more) 512 { 513 struct sk_buff *skb = q->rx_head; 514 struct skb_shared_info *shinfo = skb_shinfo(skb); 515 int nr_frags = shinfo->nr_frags; 516 517 if (nr_frags < ARRAY_SIZE(shinfo->frags)) { 518 struct page *page = virt_to_head_page(data); 519 int offset = data - page_address(page) + q->buf_offset; 520 521 skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size); 522 } else { 523 skb_free_frag(data); 524 } 525 526 if (more) 527 return; 528 529 q->rx_head = NULL; 530 if (nr_frags < ARRAY_SIZE(shinfo->frags)) 531 dev->drv->rx_skb(dev, q - dev->q_rx, skb); 532 else 533 dev_kfree_skb(skb); 534 } 535 536 static int 537 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget) 538 { 539 int len, data_len, done = 0; 540 struct sk_buff *skb; 541 unsigned char *data; 542 bool more; 543 544 while (done < budget) { 545 u32 info; 546 547 data = mt76_dma_dequeue(dev, q, false, &len, &info, &more); 548 if (!data) 549 break; 550 551 if (q->rx_head) 552 data_len = q->buf_size; 553 else 554 data_len = SKB_WITH_OVERHEAD(q->buf_size); 555 556 if (data_len < len + q->buf_offset) { 557 dev_kfree_skb(q->rx_head); 558 q->rx_head = NULL; 559 560 skb_free_frag(data); 561 continue; 562 } 563 564 if (q->rx_head) { 565 mt76_add_fragment(dev, q, data, len, more); 566 continue; 567 } 568 569 skb = build_skb(data, q->buf_size); 570 if (!skb) { 571 skb_free_frag(data); 572 continue; 573 } 574 skb_reserve(skb, q->buf_offset); 575 576 if (q == &dev->q_rx[MT_RXQ_MCU]) { 577 u32 *rxfce = (u32 *)skb->cb; 578 *rxfce = info; 579 } 580 581 __skb_put(skb, len); 582 done++; 583 584 if (more) { 585 q->rx_head = skb; 586 continue; 587 } 588 589 dev->drv->rx_skb(dev, q - dev->q_rx, skb); 590 } 591 592 mt76_dma_rx_fill(dev, q); 593 return done; 594 } 595 596 static int 597 mt76_dma_rx_poll(struct napi_struct *napi, int budget) 598 { 599 struct mt76_dev *dev; 600 int qid, done = 0, cur; 601 602 dev = container_of(napi->dev, struct mt76_dev, napi_dev); 603 qid = napi - dev->napi; 604 605 local_bh_disable(); 606 rcu_read_lock(); 607 608 do { 609 cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done); 610 mt76_rx_poll_complete(dev, qid, napi); 611 done += cur; 612 } while (cur && done < budget); 613 614 rcu_read_unlock(); 615 local_bh_enable(); 616 617 if (done < budget && napi_complete(napi)) 618 dev->drv->rx_poll_complete(dev, qid); 619 620 return done; 621 } 622 623 static int 624 mt76_dma_init(struct mt76_dev *dev) 625 { 626 int i; 627 628 init_dummy_netdev(&dev->napi_dev); 629 630 mt76_for_each_q_rx(dev, i) { 631 netif_napi_add(&dev->napi_dev, &dev->napi[i], mt76_dma_rx_poll, 632 64); 633 mt76_dma_rx_fill(dev, &dev->q_rx[i]); 634 napi_enable(&dev->napi[i]); 635 } 636 637 return 0; 638 } 639 640 static const struct mt76_queue_ops mt76_dma_ops = { 641 .init = mt76_dma_init, 642 .alloc = mt76_dma_alloc_queue, 643 .tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw, 644 .tx_queue_skb = mt76_dma_tx_queue_skb, 645 .tx_cleanup = mt76_dma_tx_cleanup, 646 .rx_reset = mt76_dma_rx_reset, 647 .kick = mt76_dma_kick_queue, 648 }; 649 650 void mt76_dma_attach(struct mt76_dev *dev) 651 { 652 dev->queue_ops = &mt76_dma_ops; 653 } 654 EXPORT_SYMBOL_GPL(mt76_dma_attach); 655 656 void mt76_dma_cleanup(struct mt76_dev *dev) 657 { 658 int i; 659 660 mt76_worker_disable(&dev->tx_worker); 661 netif_napi_del(&dev->tx_napi); 662 663 for (i = 0; i < ARRAY_SIZE(dev->phy.q_tx); i++) { 664 mt76_dma_tx_cleanup(dev, dev->phy.q_tx[i], true); 665 if (dev->phy2) 666 mt76_dma_tx_cleanup(dev, dev->phy2->q_tx[i], true); 667 } 668 669 for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++) 670 mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true); 671 672 mt76_for_each_q_rx(dev, i) { 673 netif_napi_del(&dev->napi[i]); 674 mt76_dma_rx_cleanup(dev, &dev->q_rx[i]); 675 } 676 677 mt76_free_pending_txwi(dev); 678 } 679 EXPORT_SYMBOL_GPL(mt76_dma_cleanup); 680