1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright(c) 2018 Intel Corporation. */ 3 4 #include <linux/bpf_trace.h> 5 #include <net/xdp_sock_drv.h> 6 #include <net/xdp.h> 7 8 #include "i40e.h" 9 #include "i40e_txrx_common.h" 10 #include "i40e_xsk.h" 11 12 int i40e_alloc_rx_bi_zc(struct i40e_ring *rx_ring) 13 { 14 unsigned long sz = sizeof(*rx_ring->rx_bi_zc) * rx_ring->count; 15 16 rx_ring->rx_bi_zc = kzalloc(sz, GFP_KERNEL); 17 return rx_ring->rx_bi_zc ? 0 : -ENOMEM; 18 } 19 20 void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring) 21 { 22 memset(rx_ring->rx_bi_zc, 0, 23 sizeof(*rx_ring->rx_bi_zc) * rx_ring->count); 24 } 25 26 static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx) 27 { 28 return &rx_ring->rx_bi_zc[idx]; 29 } 30 31 /** 32 * i40e_xsk_pool_enable - Enable/associate an AF_XDP buffer pool to a 33 * certain ring/qid 34 * @vsi: Current VSI 35 * @pool: buffer pool 36 * @qid: Rx ring to associate buffer pool with 37 * 38 * Returns 0 on success, <0 on failure 39 **/ 40 static int i40e_xsk_pool_enable(struct i40e_vsi *vsi, 41 struct xsk_buff_pool *pool, 42 u16 qid) 43 { 44 struct net_device *netdev = vsi->netdev; 45 bool if_running; 46 int err; 47 48 if (vsi->type != I40E_VSI_MAIN) 49 return -EINVAL; 50 51 if (qid >= vsi->num_queue_pairs) 52 return -EINVAL; 53 54 if (qid >= netdev->real_num_rx_queues || 55 qid >= netdev->real_num_tx_queues) 56 return -EINVAL; 57 58 err = xsk_pool_dma_map(pool, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR); 59 if (err) 60 return err; 61 62 set_bit(qid, vsi->af_xdp_zc_qps); 63 64 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi); 65 66 if (if_running) { 67 err = i40e_queue_pair_disable(vsi, qid); 68 if (err) 69 return err; 70 71 err = i40e_queue_pair_enable(vsi, qid); 72 if (err) 73 return err; 74 75 /* Kick start the NAPI context so that receiving will start */ 76 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX); 77 if (err) 78 return err; 79 } 80 81 return 0; 82 } 83 84 /** 85 * i40e_xsk_pool_disable - Disassociate an AF_XDP buffer pool from a 86 * certain ring/qid 87 * @vsi: Current VSI 88 * @qid: Rx ring to associate buffer pool with 89 * 90 * Returns 0 on success, <0 on failure 91 **/ 92 static int i40e_xsk_pool_disable(struct i40e_vsi *vsi, u16 qid) 93 { 94 struct net_device *netdev = vsi->netdev; 95 struct xsk_buff_pool *pool; 96 bool if_running; 97 int err; 98 99 pool = xsk_get_pool_from_qid(netdev, qid); 100 if (!pool) 101 return -EINVAL; 102 103 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi); 104 105 if (if_running) { 106 err = i40e_queue_pair_disable(vsi, qid); 107 if (err) 108 return err; 109 } 110 111 clear_bit(qid, vsi->af_xdp_zc_qps); 112 xsk_pool_dma_unmap(pool, I40E_RX_DMA_ATTR); 113 114 if (if_running) { 115 err = i40e_queue_pair_enable(vsi, qid); 116 if (err) 117 return err; 118 } 119 120 return 0; 121 } 122 123 /** 124 * i40e_xsk_pool_setup - Enable/disassociate an AF_XDP buffer pool to/from 125 * a ring/qid 126 * @vsi: Current VSI 127 * @pool: Buffer pool to enable/associate to a ring, or NULL to disable 128 * @qid: Rx ring to (dis)associate buffer pool (from)to 129 * 130 * This function enables or disables a buffer pool to a certain ring. 131 * 132 * Returns 0 on success, <0 on failure 133 **/ 134 int i40e_xsk_pool_setup(struct i40e_vsi *vsi, struct xsk_buff_pool *pool, 135 u16 qid) 136 { 137 return pool ? i40e_xsk_pool_enable(vsi, pool, qid) : 138 i40e_xsk_pool_disable(vsi, qid); 139 } 140 141 /** 142 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff 143 * @rx_ring: Rx ring 144 * @xdp: xdp_buff used as input to the XDP program 145 * 146 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR} 147 **/ 148 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp) 149 { 150 int err, result = I40E_XDP_PASS; 151 struct i40e_ring *xdp_ring; 152 struct bpf_prog *xdp_prog; 153 u32 act; 154 155 rcu_read_lock(); 156 /* NB! xdp_prog will always be !NULL, due to the fact that 157 * this path is enabled by setting an XDP program. 158 */ 159 xdp_prog = READ_ONCE(rx_ring->xdp_prog); 160 act = bpf_prog_run_xdp(xdp_prog, xdp); 161 162 switch (act) { 163 case XDP_PASS: 164 break; 165 case XDP_TX: 166 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index]; 167 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring); 168 break; 169 case XDP_REDIRECT: 170 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog); 171 result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED; 172 break; 173 default: 174 bpf_warn_invalid_xdp_action(act); 175 fallthrough; 176 case XDP_ABORTED: 177 trace_xdp_exception(rx_ring->netdev, xdp_prog, act); 178 fallthrough; /* handle aborts by dropping packet */ 179 case XDP_DROP: 180 result = I40E_XDP_CONSUMED; 181 break; 182 } 183 rcu_read_unlock(); 184 return result; 185 } 186 187 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count) 188 { 189 u16 ntu = rx_ring->next_to_use; 190 union i40e_rx_desc *rx_desc; 191 struct xdp_buff **bi, *xdp; 192 dma_addr_t dma; 193 bool ok = true; 194 195 rx_desc = I40E_RX_DESC(rx_ring, ntu); 196 bi = i40e_rx_bi(rx_ring, ntu); 197 do { 198 xdp = xsk_buff_alloc(rx_ring->xsk_pool); 199 if (!xdp) { 200 ok = false; 201 goto no_buffers; 202 } 203 *bi = xdp; 204 dma = xsk_buff_xdp_get_dma(xdp); 205 rx_desc->read.pkt_addr = cpu_to_le64(dma); 206 rx_desc->read.hdr_addr = 0; 207 208 rx_desc++; 209 bi++; 210 ntu++; 211 212 if (unlikely(ntu == rx_ring->count)) { 213 rx_desc = I40E_RX_DESC(rx_ring, 0); 214 bi = i40e_rx_bi(rx_ring, 0); 215 ntu = 0; 216 } 217 218 count--; 219 } while (count); 220 221 no_buffers: 222 if (rx_ring->next_to_use != ntu) 223 i40e_release_rx_desc(rx_ring, ntu); 224 225 return ok; 226 } 227 228 /** 229 * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer 230 * @rx_ring: Rx ring 231 * @xdp: xdp_buff 232 * 233 * This functions allocates a new skb from a zero-copy Rx buffer. 234 * 235 * Returns the skb, or NULL on failure. 236 **/ 237 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring, 238 struct xdp_buff *xdp) 239 { 240 unsigned int metasize = xdp->data - xdp->data_meta; 241 unsigned int datasize = xdp->data_end - xdp->data; 242 struct sk_buff *skb; 243 244 /* allocate a skb to store the frags */ 245 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, 246 xdp->data_end - xdp->data_hard_start, 247 GFP_ATOMIC | __GFP_NOWARN); 248 if (unlikely(!skb)) 249 return NULL; 250 251 skb_reserve(skb, xdp->data - xdp->data_hard_start); 252 memcpy(__skb_put(skb, datasize), xdp->data, datasize); 253 if (metasize) 254 skb_metadata_set(skb, metasize); 255 256 xsk_buff_free(xdp); 257 return skb; 258 } 259 260 /** 261 * i40e_inc_ntc: Advance the next_to_clean index 262 * @rx_ring: Rx ring 263 **/ 264 static void i40e_inc_ntc(struct i40e_ring *rx_ring) 265 { 266 u32 ntc = rx_ring->next_to_clean + 1; 267 268 ntc = (ntc < rx_ring->count) ? ntc : 0; 269 rx_ring->next_to_clean = ntc; 270 } 271 272 /** 273 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring 274 * @rx_ring: Rx ring 275 * @budget: NAPI budget 276 * 277 * Returns amount of work completed 278 **/ 279 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget) 280 { 281 unsigned int total_rx_bytes = 0, total_rx_packets = 0; 282 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring); 283 unsigned int xdp_res, xdp_xmit = 0; 284 struct sk_buff *skb; 285 bool failure; 286 287 while (likely(total_rx_packets < (unsigned int)budget)) { 288 union i40e_rx_desc *rx_desc; 289 struct xdp_buff **bi; 290 unsigned int size; 291 u64 qword; 292 293 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean); 294 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len); 295 296 /* This memory barrier is needed to keep us from reading 297 * any other fields out of the rx_desc until we have 298 * verified the descriptor has been written back. 299 */ 300 dma_rmb(); 301 302 if (i40e_rx_is_programming_status(qword)) { 303 i40e_clean_programming_status(rx_ring, 304 rx_desc->raw.qword[0], 305 qword); 306 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 307 xsk_buff_free(*bi); 308 *bi = NULL; 309 cleaned_count++; 310 i40e_inc_ntc(rx_ring); 311 continue; 312 } 313 314 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 315 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >> 316 I40E_RXD_QW1_LENGTH_PBUF_SHIFT; 317 if (!size) 318 break; 319 320 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 321 (*bi)->data_end = (*bi)->data + size; 322 xsk_buff_dma_sync_for_cpu(*bi, rx_ring->xsk_pool); 323 324 xdp_res = i40e_run_xdp_zc(rx_ring, *bi); 325 if (xdp_res) { 326 if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) 327 xdp_xmit |= xdp_res; 328 else 329 xsk_buff_free(*bi); 330 331 *bi = NULL; 332 total_rx_bytes += size; 333 total_rx_packets++; 334 335 cleaned_count++; 336 i40e_inc_ntc(rx_ring); 337 continue; 338 } 339 340 /* XDP_PASS path */ 341 342 /* NB! We are not checking for errors using 343 * i40e_test_staterr with 344 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that 345 * SBP is *not* set in PRT_SBPVSI (default not set). 346 */ 347 skb = i40e_construct_skb_zc(rx_ring, *bi); 348 *bi = NULL; 349 if (!skb) { 350 rx_ring->rx_stats.alloc_buff_failed++; 351 break; 352 } 353 354 cleaned_count++; 355 i40e_inc_ntc(rx_ring); 356 357 if (eth_skb_pad(skb)) 358 continue; 359 360 total_rx_bytes += skb->len; 361 total_rx_packets++; 362 363 i40e_process_skb_fields(rx_ring, rx_desc, skb); 364 napi_gro_receive(&rx_ring->q_vector->napi, skb); 365 } 366 367 if (cleaned_count >= I40E_RX_BUFFER_WRITE) 368 failure = !i40e_alloc_rx_buffers_zc(rx_ring, cleaned_count); 369 370 i40e_finalize_xdp_rx(rx_ring, xdp_xmit); 371 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets); 372 373 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) { 374 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use) 375 xsk_set_rx_need_wakeup(rx_ring->xsk_pool); 376 else 377 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool); 378 379 return (int)total_rx_packets; 380 } 381 return failure ? budget : (int)total_rx_packets; 382 } 383 384 /** 385 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP 386 * @xdp_ring: XDP Tx ring 387 * @budget: NAPI budget 388 * 389 * Returns true if the work is finished. 390 **/ 391 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) 392 { 393 unsigned int sent_frames = 0, total_bytes = 0; 394 struct i40e_tx_desc *tx_desc = NULL; 395 struct i40e_tx_buffer *tx_bi; 396 struct xdp_desc desc; 397 dma_addr_t dma; 398 399 while (budget-- > 0) { 400 if (!xsk_tx_peek_desc(xdp_ring->xsk_pool, &desc)) 401 break; 402 403 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc.addr); 404 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, 405 desc.len); 406 407 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use]; 408 tx_bi->bytecount = desc.len; 409 410 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use); 411 tx_desc->buffer_addr = cpu_to_le64(dma); 412 tx_desc->cmd_type_offset_bsz = 413 build_ctob(I40E_TX_DESC_CMD_ICRC 414 | I40E_TX_DESC_CMD_EOP, 415 0, desc.len, 0); 416 417 sent_frames++; 418 total_bytes += tx_bi->bytecount; 419 420 xdp_ring->next_to_use++; 421 if (xdp_ring->next_to_use == xdp_ring->count) 422 xdp_ring->next_to_use = 0; 423 } 424 425 if (tx_desc) { 426 /* Request an interrupt for the last frame and bump tail ptr. */ 427 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS << 428 I40E_TXD_QW1_CMD_SHIFT); 429 i40e_xdp_ring_update_tail(xdp_ring); 430 431 xsk_tx_release(xdp_ring->xsk_pool); 432 i40e_update_tx_stats(xdp_ring, sent_frames, total_bytes); 433 } 434 435 return !!budget; 436 } 437 438 /** 439 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry 440 * @tx_ring: XDP Tx ring 441 * @tx_bi: Tx buffer info to clean 442 **/ 443 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring, 444 struct i40e_tx_buffer *tx_bi) 445 { 446 xdp_return_frame(tx_bi->xdpf); 447 tx_ring->xdp_tx_active--; 448 dma_unmap_single(tx_ring->dev, 449 dma_unmap_addr(tx_bi, dma), 450 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE); 451 dma_unmap_len_set(tx_bi, len, 0); 452 } 453 454 /** 455 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries 456 * @vsi: Current VSI 457 * @tx_ring: XDP Tx ring 458 * 459 * Returns true if cleanup/tranmission is done. 460 **/ 461 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring) 462 { 463 struct xsk_buff_pool *bp = tx_ring->xsk_pool; 464 u32 i, completed_frames, xsk_frames = 0; 465 u32 head_idx = i40e_get_head(tx_ring); 466 struct i40e_tx_buffer *tx_bi; 467 unsigned int ntc; 468 469 if (head_idx < tx_ring->next_to_clean) 470 head_idx += tx_ring->count; 471 completed_frames = head_idx - tx_ring->next_to_clean; 472 473 if (completed_frames == 0) 474 goto out_xmit; 475 476 if (likely(!tx_ring->xdp_tx_active)) { 477 xsk_frames = completed_frames; 478 goto skip; 479 } 480 481 ntc = tx_ring->next_to_clean; 482 483 for (i = 0; i < completed_frames; i++) { 484 tx_bi = &tx_ring->tx_bi[ntc]; 485 486 if (tx_bi->xdpf) { 487 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi); 488 tx_bi->xdpf = NULL; 489 } else { 490 xsk_frames++; 491 } 492 493 if (++ntc >= tx_ring->count) 494 ntc = 0; 495 } 496 497 skip: 498 tx_ring->next_to_clean += completed_frames; 499 if (unlikely(tx_ring->next_to_clean >= tx_ring->count)) 500 tx_ring->next_to_clean -= tx_ring->count; 501 502 if (xsk_frames) 503 xsk_tx_completed(bp, xsk_frames); 504 505 i40e_arm_wb(tx_ring, vsi, completed_frames); 506 507 out_xmit: 508 if (xsk_uses_need_wakeup(tx_ring->xsk_pool)) 509 xsk_set_tx_need_wakeup(tx_ring->xsk_pool); 510 511 return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring)); 512 } 513 514 /** 515 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup 516 * @dev: the netdevice 517 * @queue_id: queue id to wake up 518 * @flags: ignored in our case since we have Rx and Tx in the same NAPI. 519 * 520 * Returns <0 for errors, 0 otherwise. 521 **/ 522 int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags) 523 { 524 struct i40e_netdev_priv *np = netdev_priv(dev); 525 struct i40e_vsi *vsi = np->vsi; 526 struct i40e_pf *pf = vsi->back; 527 struct i40e_ring *ring; 528 529 if (test_bit(__I40E_CONFIG_BUSY, pf->state)) 530 return -EAGAIN; 531 532 if (test_bit(__I40E_VSI_DOWN, vsi->state)) 533 return -ENETDOWN; 534 535 if (!i40e_enabled_xdp_vsi(vsi)) 536 return -ENXIO; 537 538 if (queue_id >= vsi->num_queue_pairs) 539 return -ENXIO; 540 541 if (!vsi->xdp_rings[queue_id]->xsk_pool) 542 return -ENXIO; 543 544 ring = vsi->xdp_rings[queue_id]; 545 546 /* The idea here is that if NAPI is running, mark a miss, so 547 * it will run again. If not, trigger an interrupt and 548 * schedule the NAPI from interrupt context. If NAPI would be 549 * scheduled here, the interrupt affinity would not be 550 * honored. 551 */ 552 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi)) 553 i40e_force_wb(vsi, ring->q_vector); 554 555 return 0; 556 } 557 558 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring) 559 { 560 u16 i; 561 562 for (i = 0; i < rx_ring->count; i++) { 563 struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, i); 564 565 if (!rx_bi) 566 continue; 567 568 xsk_buff_free(rx_bi); 569 rx_bi = NULL; 570 } 571 } 572 573 /** 574 * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown 575 * @tx_ring: XDP Tx ring 576 **/ 577 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring) 578 { 579 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 580 struct xsk_buff_pool *bp = tx_ring->xsk_pool; 581 struct i40e_tx_buffer *tx_bi; 582 u32 xsk_frames = 0; 583 584 while (ntc != ntu) { 585 tx_bi = &tx_ring->tx_bi[ntc]; 586 587 if (tx_bi->xdpf) 588 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi); 589 else 590 xsk_frames++; 591 592 tx_bi->xdpf = NULL; 593 594 ntc++; 595 if (ntc >= tx_ring->count) 596 ntc = 0; 597 } 598 599 if (xsk_frames) 600 xsk_tx_completed(bp, xsk_frames); 601 } 602 603 /** 604 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP 605 * buffer pool attached 606 * @vsi: vsi 607 * 608 * Returns true if any of the Rx rings has an AF_XDP buffer pool attached 609 **/ 610 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi) 611 { 612 struct net_device *netdev = vsi->netdev; 613 int i; 614 615 for (i = 0; i < vsi->num_queue_pairs; i++) { 616 if (xsk_get_pool_from_qid(netdev, i)) 617 return true; 618 } 619 620 return false; 621 } 622