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_clean_rx_irq_zc - Consumes Rx packets from the hardware ring 262 * @rx_ring: Rx ring 263 * @budget: NAPI budget 264 * 265 * Returns amount of work completed 266 **/ 267 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget) 268 { 269 unsigned int total_rx_bytes = 0, total_rx_packets = 0; 270 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring); 271 unsigned int xdp_res, xdp_xmit = 0; 272 bool failure = false; 273 struct sk_buff *skb; 274 275 while (likely(total_rx_packets < (unsigned int)budget)) { 276 union i40e_rx_desc *rx_desc; 277 struct xdp_buff **bi; 278 unsigned int size; 279 u64 qword; 280 281 if (cleaned_count >= I40E_RX_BUFFER_WRITE) { 282 failure = failure || 283 !i40e_alloc_rx_buffers_zc(rx_ring, 284 cleaned_count); 285 cleaned_count = 0; 286 } 287 288 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean); 289 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len); 290 291 /* This memory barrier is needed to keep us from reading 292 * any other fields out of the rx_desc until we have 293 * verified the descriptor has been written back. 294 */ 295 dma_rmb(); 296 297 if (i40e_rx_is_programming_status(qword)) { 298 i40e_clean_programming_status(rx_ring, 299 rx_desc->raw.qword[0], 300 qword); 301 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 302 xsk_buff_free(*bi); 303 *bi = NULL; 304 cleaned_count++; 305 i40e_inc_ntc(rx_ring); 306 continue; 307 } 308 309 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 310 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >> 311 I40E_RXD_QW1_LENGTH_PBUF_SHIFT; 312 if (!size) 313 break; 314 315 bi = i40e_rx_bi(rx_ring, rx_ring->next_to_clean); 316 (*bi)->data_end = (*bi)->data + size; 317 xsk_buff_dma_sync_for_cpu(*bi, rx_ring->xsk_pool); 318 319 xdp_res = i40e_run_xdp_zc(rx_ring, *bi); 320 if (xdp_res) { 321 if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) 322 xdp_xmit |= xdp_res; 323 else 324 xsk_buff_free(*bi); 325 326 *bi = NULL; 327 total_rx_bytes += size; 328 total_rx_packets++; 329 330 cleaned_count++; 331 i40e_inc_ntc(rx_ring); 332 continue; 333 } 334 335 /* XDP_PASS path */ 336 337 /* NB! We are not checking for errors using 338 * i40e_test_staterr with 339 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that 340 * SBP is *not* set in PRT_SBPVSI (default not set). 341 */ 342 skb = i40e_construct_skb_zc(rx_ring, *bi); 343 *bi = NULL; 344 if (!skb) { 345 rx_ring->rx_stats.alloc_buff_failed++; 346 break; 347 } 348 349 cleaned_count++; 350 i40e_inc_ntc(rx_ring); 351 352 if (eth_skb_pad(skb)) 353 continue; 354 355 total_rx_bytes += skb->len; 356 total_rx_packets++; 357 358 i40e_process_skb_fields(rx_ring, rx_desc, skb); 359 napi_gro_receive(&rx_ring->q_vector->napi, skb); 360 } 361 362 i40e_finalize_xdp_rx(rx_ring, xdp_xmit); 363 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets); 364 365 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) { 366 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use) 367 xsk_set_rx_need_wakeup(rx_ring->xsk_pool); 368 else 369 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool); 370 371 return (int)total_rx_packets; 372 } 373 return failure ? budget : (int)total_rx_packets; 374 } 375 376 /** 377 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP 378 * @xdp_ring: XDP Tx ring 379 * @budget: NAPI budget 380 * 381 * Returns true if the work is finished. 382 **/ 383 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget) 384 { 385 unsigned int sent_frames = 0, total_bytes = 0; 386 struct i40e_tx_desc *tx_desc = NULL; 387 struct i40e_tx_buffer *tx_bi; 388 struct xdp_desc desc; 389 dma_addr_t dma; 390 391 while (budget-- > 0) { 392 if (!xsk_tx_peek_desc(xdp_ring->xsk_pool, &desc)) 393 break; 394 395 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc.addr); 396 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, 397 desc.len); 398 399 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use]; 400 tx_bi->bytecount = desc.len; 401 402 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use); 403 tx_desc->buffer_addr = cpu_to_le64(dma); 404 tx_desc->cmd_type_offset_bsz = 405 build_ctob(I40E_TX_DESC_CMD_ICRC 406 | I40E_TX_DESC_CMD_EOP, 407 0, desc.len, 0); 408 409 sent_frames++; 410 total_bytes += tx_bi->bytecount; 411 412 xdp_ring->next_to_use++; 413 if (xdp_ring->next_to_use == xdp_ring->count) 414 xdp_ring->next_to_use = 0; 415 } 416 417 if (tx_desc) { 418 /* Request an interrupt for the last frame and bump tail ptr. */ 419 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS << 420 I40E_TXD_QW1_CMD_SHIFT); 421 i40e_xdp_ring_update_tail(xdp_ring); 422 423 xsk_tx_release(xdp_ring->xsk_pool); 424 i40e_update_tx_stats(xdp_ring, sent_frames, total_bytes); 425 } 426 427 return !!budget; 428 } 429 430 /** 431 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry 432 * @tx_ring: XDP Tx ring 433 * @tx_bi: Tx buffer info to clean 434 **/ 435 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring, 436 struct i40e_tx_buffer *tx_bi) 437 { 438 xdp_return_frame(tx_bi->xdpf); 439 tx_ring->xdp_tx_active--; 440 dma_unmap_single(tx_ring->dev, 441 dma_unmap_addr(tx_bi, dma), 442 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE); 443 dma_unmap_len_set(tx_bi, len, 0); 444 } 445 446 /** 447 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries 448 * @vsi: Current VSI 449 * @tx_ring: XDP Tx ring 450 * 451 * Returns true if cleanup/tranmission is done. 452 **/ 453 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring) 454 { 455 struct xsk_buff_pool *bp = tx_ring->xsk_pool; 456 u32 i, completed_frames, xsk_frames = 0; 457 u32 head_idx = i40e_get_head(tx_ring); 458 struct i40e_tx_buffer *tx_bi; 459 unsigned int ntc; 460 461 if (head_idx < tx_ring->next_to_clean) 462 head_idx += tx_ring->count; 463 completed_frames = head_idx - tx_ring->next_to_clean; 464 465 if (completed_frames == 0) 466 goto out_xmit; 467 468 if (likely(!tx_ring->xdp_tx_active)) { 469 xsk_frames = completed_frames; 470 goto skip; 471 } 472 473 ntc = tx_ring->next_to_clean; 474 475 for (i = 0; i < completed_frames; i++) { 476 tx_bi = &tx_ring->tx_bi[ntc]; 477 478 if (tx_bi->xdpf) { 479 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi); 480 tx_bi->xdpf = NULL; 481 } else { 482 xsk_frames++; 483 } 484 485 if (++ntc >= tx_ring->count) 486 ntc = 0; 487 } 488 489 skip: 490 tx_ring->next_to_clean += completed_frames; 491 if (unlikely(tx_ring->next_to_clean >= tx_ring->count)) 492 tx_ring->next_to_clean -= tx_ring->count; 493 494 if (xsk_frames) 495 xsk_tx_completed(bp, xsk_frames); 496 497 i40e_arm_wb(tx_ring, vsi, completed_frames); 498 499 out_xmit: 500 if (xsk_uses_need_wakeup(tx_ring->xsk_pool)) 501 xsk_set_tx_need_wakeup(tx_ring->xsk_pool); 502 503 return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring)); 504 } 505 506 /** 507 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup 508 * @dev: the netdevice 509 * @queue_id: queue id to wake up 510 * @flags: ignored in our case since we have Rx and Tx in the same NAPI. 511 * 512 * Returns <0 for errors, 0 otherwise. 513 **/ 514 int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags) 515 { 516 struct i40e_netdev_priv *np = netdev_priv(dev); 517 struct i40e_vsi *vsi = np->vsi; 518 struct i40e_pf *pf = vsi->back; 519 struct i40e_ring *ring; 520 521 if (test_bit(__I40E_CONFIG_BUSY, pf->state)) 522 return -EAGAIN; 523 524 if (test_bit(__I40E_VSI_DOWN, vsi->state)) 525 return -ENETDOWN; 526 527 if (!i40e_enabled_xdp_vsi(vsi)) 528 return -ENXIO; 529 530 if (queue_id >= vsi->num_queue_pairs) 531 return -ENXIO; 532 533 if (!vsi->xdp_rings[queue_id]->xsk_pool) 534 return -ENXIO; 535 536 ring = vsi->xdp_rings[queue_id]; 537 538 /* The idea here is that if NAPI is running, mark a miss, so 539 * it will run again. If not, trigger an interrupt and 540 * schedule the NAPI from interrupt context. If NAPI would be 541 * scheduled here, the interrupt affinity would not be 542 * honored. 543 */ 544 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi)) 545 i40e_force_wb(vsi, ring->q_vector); 546 547 return 0; 548 } 549 550 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring) 551 { 552 u16 i; 553 554 for (i = 0; i < rx_ring->count; i++) { 555 struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, i); 556 557 if (!rx_bi) 558 continue; 559 560 xsk_buff_free(rx_bi); 561 rx_bi = NULL; 562 } 563 } 564 565 /** 566 * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown 567 * @tx_ring: XDP Tx ring 568 **/ 569 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring) 570 { 571 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 572 struct xsk_buff_pool *bp = tx_ring->xsk_pool; 573 struct i40e_tx_buffer *tx_bi; 574 u32 xsk_frames = 0; 575 576 while (ntc != ntu) { 577 tx_bi = &tx_ring->tx_bi[ntc]; 578 579 if (tx_bi->xdpf) 580 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi); 581 else 582 xsk_frames++; 583 584 tx_bi->xdpf = NULL; 585 586 ntc++; 587 if (ntc >= tx_ring->count) 588 ntc = 0; 589 } 590 591 if (xsk_frames) 592 xsk_tx_completed(bp, xsk_frames); 593 } 594 595 /** 596 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP 597 * buffer pool attached 598 * @vsi: vsi 599 * 600 * Returns true if any of the Rx rings has an AF_XDP buffer pool attached 601 **/ 602 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi) 603 { 604 struct net_device *netdev = vsi->netdev; 605 int i; 606 607 for (i = 0; i < vsi->num_queue_pairs; i++) { 608 if (xsk_get_pool_from_qid(netdev, i)) 609 return true; 610 } 611 612 return false; 613 } 614