1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2019, Intel Corporation. */ 3 4 #include <linux/bpf_trace.h> 5 #include <net/xdp_sock_drv.h> 6 #include <net/xdp.h> 7 #include "ice.h" 8 #include "ice_base.h" 9 #include "ice_type.h" 10 #include "ice_xsk.h" 11 #include "ice_txrx.h" 12 #include "ice_txrx_lib.h" 13 #include "ice_lib.h" 14 15 /** 16 * ice_qp_reset_stats - Resets all stats for rings of given index 17 * @vsi: VSI that contains rings of interest 18 * @q_idx: ring index in array 19 */ 20 static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx) 21 { 22 memset(&vsi->rx_rings[q_idx]->rx_stats, 0, 23 sizeof(vsi->rx_rings[q_idx]->rx_stats)); 24 memset(&vsi->tx_rings[q_idx]->stats, 0, 25 sizeof(vsi->tx_rings[q_idx]->stats)); 26 if (ice_is_xdp_ena_vsi(vsi)) 27 memset(&vsi->xdp_rings[q_idx]->stats, 0, 28 sizeof(vsi->xdp_rings[q_idx]->stats)); 29 } 30 31 /** 32 * ice_qp_clean_rings - Cleans all the rings of a given index 33 * @vsi: VSI that contains rings of interest 34 * @q_idx: ring index in array 35 */ 36 static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx) 37 { 38 ice_clean_tx_ring(vsi->tx_rings[q_idx]); 39 if (ice_is_xdp_ena_vsi(vsi)) 40 ice_clean_tx_ring(vsi->xdp_rings[q_idx]); 41 ice_clean_rx_ring(vsi->rx_rings[q_idx]); 42 } 43 44 /** 45 * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector 46 * @vsi: VSI that has netdev 47 * @q_vector: q_vector that has NAPI context 48 * @enable: true for enable, false for disable 49 */ 50 static void 51 ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector, 52 bool enable) 53 { 54 if (!vsi->netdev || !q_vector) 55 return; 56 57 if (enable) 58 napi_enable(&q_vector->napi); 59 else 60 napi_disable(&q_vector->napi); 61 } 62 63 /** 64 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring 65 * @vsi: the VSI that contains queue vector being un-configured 66 * @rx_ring: Rx ring that will have its IRQ disabled 67 * @q_vector: queue vector 68 */ 69 static void 70 ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_ring *rx_ring, 71 struct ice_q_vector *q_vector) 72 { 73 struct ice_pf *pf = vsi->back; 74 struct ice_hw *hw = &pf->hw; 75 int base = vsi->base_vector; 76 u16 reg; 77 u32 val; 78 79 /* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle 80 * here only QINT_RQCTL 81 */ 82 reg = rx_ring->reg_idx; 83 val = rd32(hw, QINT_RQCTL(reg)); 84 val &= ~QINT_RQCTL_CAUSE_ENA_M; 85 wr32(hw, QINT_RQCTL(reg), val); 86 87 if (q_vector) { 88 u16 v_idx = q_vector->v_idx; 89 90 wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0); 91 ice_flush(hw); 92 synchronize_irq(pf->msix_entries[v_idx + base].vector); 93 } 94 } 95 96 /** 97 * ice_qvec_cfg_msix - Enable IRQ for given queue vector 98 * @vsi: the VSI that contains queue vector 99 * @q_vector: queue vector 100 */ 101 static void 102 ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector) 103 { 104 u16 reg_idx = q_vector->reg_idx; 105 struct ice_pf *pf = vsi->back; 106 struct ice_hw *hw = &pf->hw; 107 struct ice_ring *ring; 108 109 ice_cfg_itr(hw, q_vector); 110 111 wr32(hw, GLINT_RATE(reg_idx), 112 ice_intrl_usec_to_reg(q_vector->intrl, hw->intrl_gran)); 113 114 ice_for_each_ring(ring, q_vector->tx) 115 ice_cfg_txq_interrupt(vsi, ring->reg_idx, reg_idx, 116 q_vector->tx.itr_idx); 117 118 ice_for_each_ring(ring, q_vector->rx) 119 ice_cfg_rxq_interrupt(vsi, ring->reg_idx, reg_idx, 120 q_vector->rx.itr_idx); 121 122 ice_flush(hw); 123 } 124 125 /** 126 * ice_qvec_ena_irq - Enable IRQ for given queue vector 127 * @vsi: the VSI that contains queue vector 128 * @q_vector: queue vector 129 */ 130 static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector) 131 { 132 struct ice_pf *pf = vsi->back; 133 struct ice_hw *hw = &pf->hw; 134 135 ice_irq_dynamic_ena(hw, vsi, q_vector); 136 137 ice_flush(hw); 138 } 139 140 /** 141 * ice_qp_dis - Disables a queue pair 142 * @vsi: VSI of interest 143 * @q_idx: ring index in array 144 * 145 * Returns 0 on success, negative on failure. 146 */ 147 static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx) 148 { 149 struct ice_txq_meta txq_meta = { }; 150 struct ice_ring *tx_ring, *rx_ring; 151 struct ice_q_vector *q_vector; 152 int timeout = 50; 153 int err; 154 155 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq) 156 return -EINVAL; 157 158 tx_ring = vsi->tx_rings[q_idx]; 159 rx_ring = vsi->rx_rings[q_idx]; 160 q_vector = rx_ring->q_vector; 161 162 while (test_and_set_bit(__ICE_CFG_BUSY, vsi->state)) { 163 timeout--; 164 if (!timeout) 165 return -EBUSY; 166 usleep_range(1000, 2000); 167 } 168 netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx)); 169 170 ice_qvec_dis_irq(vsi, rx_ring, q_vector); 171 172 ice_fill_txq_meta(vsi, tx_ring, &txq_meta); 173 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta); 174 if (err) 175 return err; 176 if (ice_is_xdp_ena_vsi(vsi)) { 177 struct ice_ring *xdp_ring = vsi->xdp_rings[q_idx]; 178 179 memset(&txq_meta, 0, sizeof(txq_meta)); 180 ice_fill_txq_meta(vsi, xdp_ring, &txq_meta); 181 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring, 182 &txq_meta); 183 if (err) 184 return err; 185 } 186 err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true); 187 if (err) 188 return err; 189 190 ice_qvec_toggle_napi(vsi, q_vector, false); 191 ice_qp_clean_rings(vsi, q_idx); 192 ice_qp_reset_stats(vsi, q_idx); 193 194 return 0; 195 } 196 197 /** 198 * ice_qp_ena - Enables a queue pair 199 * @vsi: VSI of interest 200 * @q_idx: ring index in array 201 * 202 * Returns 0 on success, negative on failure. 203 */ 204 static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx) 205 { 206 struct ice_aqc_add_tx_qgrp *qg_buf; 207 struct ice_ring *tx_ring, *rx_ring; 208 struct ice_q_vector *q_vector; 209 u16 size; 210 int err; 211 212 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq) 213 return -EINVAL; 214 215 size = struct_size(qg_buf, txqs, 1); 216 qg_buf = kzalloc(size, GFP_KERNEL); 217 if (!qg_buf) 218 return -ENOMEM; 219 220 qg_buf->num_txqs = 1; 221 222 tx_ring = vsi->tx_rings[q_idx]; 223 rx_ring = vsi->rx_rings[q_idx]; 224 q_vector = rx_ring->q_vector; 225 226 err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf); 227 if (err) 228 goto free_buf; 229 230 if (ice_is_xdp_ena_vsi(vsi)) { 231 struct ice_ring *xdp_ring = vsi->xdp_rings[q_idx]; 232 233 memset(qg_buf, 0, size); 234 qg_buf->num_txqs = 1; 235 err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf); 236 if (err) 237 goto free_buf; 238 ice_set_ring_xdp(xdp_ring); 239 xdp_ring->xsk_umem = ice_xsk_umem(xdp_ring); 240 } 241 242 err = ice_setup_rx_ctx(rx_ring); 243 if (err) 244 goto free_buf; 245 246 ice_qvec_cfg_msix(vsi, q_vector); 247 248 err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true); 249 if (err) 250 goto free_buf; 251 252 clear_bit(__ICE_CFG_BUSY, vsi->state); 253 ice_qvec_toggle_napi(vsi, q_vector, true); 254 ice_qvec_ena_irq(vsi, q_vector); 255 256 netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx)); 257 free_buf: 258 kfree(qg_buf); 259 return err; 260 } 261 262 /** 263 * ice_xsk_alloc_umems - allocate a UMEM region for an XDP socket 264 * @vsi: VSI to allocate the UMEM on 265 * 266 * Returns 0 on success, negative on error 267 */ 268 static int ice_xsk_alloc_umems(struct ice_vsi *vsi) 269 { 270 if (vsi->xsk_umems) 271 return 0; 272 273 vsi->xsk_umems = kcalloc(vsi->num_xsk_umems, sizeof(*vsi->xsk_umems), 274 GFP_KERNEL); 275 276 if (!vsi->xsk_umems) { 277 vsi->num_xsk_umems = 0; 278 return -ENOMEM; 279 } 280 281 return 0; 282 } 283 284 /** 285 * ice_xsk_remove_umem - Remove an UMEM for a certain ring/qid 286 * @vsi: VSI from which the VSI will be removed 287 * @qid: Ring/qid associated with the UMEM 288 */ 289 static void ice_xsk_remove_umem(struct ice_vsi *vsi, u16 qid) 290 { 291 vsi->xsk_umems[qid] = NULL; 292 vsi->num_xsk_umems_used--; 293 294 if (vsi->num_xsk_umems_used == 0) { 295 kfree(vsi->xsk_umems); 296 vsi->xsk_umems = NULL; 297 vsi->num_xsk_umems = 0; 298 } 299 } 300 301 302 /** 303 * ice_xsk_umem_disable - disable a UMEM region 304 * @vsi: Current VSI 305 * @qid: queue ID 306 * 307 * Returns 0 on success, negative on failure 308 */ 309 static int ice_xsk_umem_disable(struct ice_vsi *vsi, u16 qid) 310 { 311 if (!vsi->xsk_umems || qid >= vsi->num_xsk_umems || 312 !vsi->xsk_umems[qid]) 313 return -EINVAL; 314 315 xsk_buff_dma_unmap(vsi->xsk_umems[qid], ICE_RX_DMA_ATTR); 316 ice_xsk_remove_umem(vsi, qid); 317 318 return 0; 319 } 320 321 /** 322 * ice_xsk_umem_enable - enable a UMEM region 323 * @vsi: Current VSI 324 * @umem: pointer to a requested UMEM region 325 * @qid: queue ID 326 * 327 * Returns 0 on success, negative on failure 328 */ 329 static int 330 ice_xsk_umem_enable(struct ice_vsi *vsi, struct xdp_umem *umem, u16 qid) 331 { 332 int err; 333 334 if (vsi->type != ICE_VSI_PF) 335 return -EINVAL; 336 337 if (!vsi->num_xsk_umems) 338 vsi->num_xsk_umems = min_t(u16, vsi->num_rxq, vsi->num_txq); 339 if (qid >= vsi->num_xsk_umems) 340 return -EINVAL; 341 342 err = ice_xsk_alloc_umems(vsi); 343 if (err) 344 return err; 345 346 if (vsi->xsk_umems && vsi->xsk_umems[qid]) 347 return -EBUSY; 348 349 vsi->xsk_umems[qid] = umem; 350 vsi->num_xsk_umems_used++; 351 352 err = xsk_buff_dma_map(vsi->xsk_umems[qid], ice_pf_to_dev(vsi->back), 353 ICE_RX_DMA_ATTR); 354 if (err) 355 return err; 356 357 return 0; 358 } 359 360 /** 361 * ice_xsk_umem_setup - enable/disable a UMEM region depending on its state 362 * @vsi: Current VSI 363 * @umem: UMEM to enable/associate to a ring, NULL to disable 364 * @qid: queue ID 365 * 366 * Returns 0 on success, negative on failure 367 */ 368 int ice_xsk_umem_setup(struct ice_vsi *vsi, struct xdp_umem *umem, u16 qid) 369 { 370 bool if_running, umem_present = !!umem; 371 int ret = 0, umem_failure = 0; 372 373 if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi); 374 375 if (if_running) { 376 ret = ice_qp_dis(vsi, qid); 377 if (ret) { 378 netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret); 379 goto xsk_umem_if_up; 380 } 381 } 382 383 umem_failure = umem_present ? ice_xsk_umem_enable(vsi, umem, qid) : 384 ice_xsk_umem_disable(vsi, qid); 385 386 xsk_umem_if_up: 387 if (if_running) { 388 ret = ice_qp_ena(vsi, qid); 389 if (!ret && umem_present) 390 napi_schedule(&vsi->xdp_rings[qid]->q_vector->napi); 391 else if (ret) 392 netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret); 393 } 394 395 if (umem_failure) { 396 netdev_err(vsi->netdev, "Could not %sable UMEM, error = %d\n", 397 umem_present ? "en" : "dis", umem_failure); 398 return umem_failure; 399 } 400 401 return ret; 402 } 403 404 /** 405 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers 406 * @rx_ring: Rx ring 407 * @count: The number of buffers to allocate 408 * 409 * This function allocates a number of Rx buffers from the fill ring 410 * or the internal recycle mechanism and places them on the Rx ring. 411 * 412 * Returns false if all allocations were successful, true if any fail. 413 */ 414 bool ice_alloc_rx_bufs_zc(struct ice_ring *rx_ring, u16 count) 415 { 416 union ice_32b_rx_flex_desc *rx_desc; 417 u16 ntu = rx_ring->next_to_use; 418 struct ice_rx_buf *rx_buf; 419 bool ret = false; 420 dma_addr_t dma; 421 422 if (!count) 423 return false; 424 425 rx_desc = ICE_RX_DESC(rx_ring, ntu); 426 rx_buf = &rx_ring->rx_buf[ntu]; 427 428 do { 429 rx_buf->xdp = xsk_buff_alloc(rx_ring->xsk_umem); 430 if (!rx_buf->xdp) { 431 ret = true; 432 break; 433 } 434 435 dma = xsk_buff_xdp_get_dma(rx_buf->xdp); 436 rx_desc->read.pkt_addr = cpu_to_le64(dma); 437 rx_desc->wb.status_error0 = 0; 438 439 rx_desc++; 440 rx_buf++; 441 ntu++; 442 443 if (unlikely(ntu == rx_ring->count)) { 444 rx_desc = ICE_RX_DESC(rx_ring, 0); 445 rx_buf = rx_ring->rx_buf; 446 ntu = 0; 447 } 448 } while (--count); 449 450 if (rx_ring->next_to_use != ntu) 451 ice_release_rx_desc(rx_ring, ntu); 452 453 return ret; 454 } 455 456 /** 457 * ice_bump_ntc - Bump the next_to_clean counter of an Rx ring 458 * @rx_ring: Rx ring 459 */ 460 static void ice_bump_ntc(struct ice_ring *rx_ring) 461 { 462 int ntc = rx_ring->next_to_clean + 1; 463 464 ntc = (ntc < rx_ring->count) ? ntc : 0; 465 rx_ring->next_to_clean = ntc; 466 prefetch(ICE_RX_DESC(rx_ring, ntc)); 467 } 468 469 /** 470 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer 471 * @rx_ring: Rx ring 472 * @rx_buf: zero-copy Rx buffer 473 * 474 * This function allocates a new skb from a zero-copy Rx buffer. 475 * 476 * Returns the skb on success, NULL on failure. 477 */ 478 static struct sk_buff * 479 ice_construct_skb_zc(struct ice_ring *rx_ring, struct ice_rx_buf *rx_buf) 480 { 481 unsigned int metasize = rx_buf->xdp->data - rx_buf->xdp->data_meta; 482 unsigned int datasize = rx_buf->xdp->data_end - rx_buf->xdp->data; 483 unsigned int datasize_hard = rx_buf->xdp->data_end - 484 rx_buf->xdp->data_hard_start; 485 struct sk_buff *skb; 486 487 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, datasize_hard, 488 GFP_ATOMIC | __GFP_NOWARN); 489 if (unlikely(!skb)) 490 return NULL; 491 492 skb_reserve(skb, rx_buf->xdp->data - rx_buf->xdp->data_hard_start); 493 memcpy(__skb_put(skb, datasize), rx_buf->xdp->data, datasize); 494 if (metasize) 495 skb_metadata_set(skb, metasize); 496 497 xsk_buff_free(rx_buf->xdp); 498 rx_buf->xdp = NULL; 499 return skb; 500 } 501 502 /** 503 * ice_run_xdp_zc - Executes an XDP program in zero-copy path 504 * @rx_ring: Rx ring 505 * @xdp: xdp_buff used as input to the XDP program 506 * 507 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR} 508 */ 509 static int 510 ice_run_xdp_zc(struct ice_ring *rx_ring, struct xdp_buff *xdp) 511 { 512 int err, result = ICE_XDP_PASS; 513 struct bpf_prog *xdp_prog; 514 struct ice_ring *xdp_ring; 515 u32 act; 516 517 rcu_read_lock(); 518 xdp_prog = READ_ONCE(rx_ring->xdp_prog); 519 if (!xdp_prog) { 520 rcu_read_unlock(); 521 return ICE_XDP_PASS; 522 } 523 524 act = bpf_prog_run_xdp(xdp_prog, xdp); 525 switch (act) { 526 case XDP_PASS: 527 break; 528 case XDP_TX: 529 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->q_index]; 530 result = ice_xmit_xdp_buff(xdp, xdp_ring); 531 break; 532 case XDP_REDIRECT: 533 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog); 534 result = !err ? ICE_XDP_REDIR : ICE_XDP_CONSUMED; 535 break; 536 default: 537 bpf_warn_invalid_xdp_action(act); 538 fallthrough; 539 case XDP_ABORTED: 540 trace_xdp_exception(rx_ring->netdev, xdp_prog, act); 541 fallthrough; 542 case XDP_DROP: 543 result = ICE_XDP_CONSUMED; 544 break; 545 } 546 547 rcu_read_unlock(); 548 return result; 549 } 550 551 /** 552 * ice_clean_rx_irq_zc - consumes packets from the hardware ring 553 * @rx_ring: AF_XDP Rx ring 554 * @budget: NAPI budget 555 * 556 * Returns number of processed packets on success, remaining budget on failure. 557 */ 558 int ice_clean_rx_irq_zc(struct ice_ring *rx_ring, int budget) 559 { 560 unsigned int total_rx_bytes = 0, total_rx_packets = 0; 561 u16 cleaned_count = ICE_DESC_UNUSED(rx_ring); 562 unsigned int xdp_xmit = 0; 563 bool failure = false; 564 565 while (likely(total_rx_packets < (unsigned int)budget)) { 566 union ice_32b_rx_flex_desc *rx_desc; 567 unsigned int size, xdp_res = 0; 568 struct ice_rx_buf *rx_buf; 569 struct sk_buff *skb; 570 u16 stat_err_bits; 571 u16 vlan_tag = 0; 572 u8 rx_ptype; 573 574 if (cleaned_count >= ICE_RX_BUF_WRITE) { 575 failure |= ice_alloc_rx_bufs_zc(rx_ring, 576 cleaned_count); 577 cleaned_count = 0; 578 } 579 580 rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean); 581 582 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S); 583 if (!ice_test_staterr(rx_desc, stat_err_bits)) 584 break; 585 586 /* This memory barrier is needed to keep us from reading 587 * any other fields out of the rx_desc until we have 588 * verified the descriptor has been written back. 589 */ 590 dma_rmb(); 591 592 size = le16_to_cpu(rx_desc->wb.pkt_len) & 593 ICE_RX_FLX_DESC_PKT_LEN_M; 594 if (!size) 595 break; 596 597 598 rx_buf = &rx_ring->rx_buf[rx_ring->next_to_clean]; 599 rx_buf->xdp->data_end = rx_buf->xdp->data + size; 600 xsk_buff_dma_sync_for_cpu(rx_buf->xdp); 601 602 xdp_res = ice_run_xdp_zc(rx_ring, rx_buf->xdp); 603 if (xdp_res) { 604 if (xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR)) 605 xdp_xmit |= xdp_res; 606 else 607 xsk_buff_free(rx_buf->xdp); 608 609 rx_buf->xdp = NULL; 610 total_rx_bytes += size; 611 total_rx_packets++; 612 cleaned_count++; 613 614 ice_bump_ntc(rx_ring); 615 continue; 616 } 617 618 /* XDP_PASS path */ 619 skb = ice_construct_skb_zc(rx_ring, rx_buf); 620 if (!skb) { 621 rx_ring->rx_stats.alloc_buf_failed++; 622 break; 623 } 624 625 cleaned_count++; 626 ice_bump_ntc(rx_ring); 627 628 if (eth_skb_pad(skb)) { 629 skb = NULL; 630 continue; 631 } 632 633 total_rx_bytes += skb->len; 634 total_rx_packets++; 635 636 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_L2TAG1P_S); 637 if (ice_test_staterr(rx_desc, stat_err_bits)) 638 vlan_tag = le16_to_cpu(rx_desc->wb.l2tag1); 639 640 rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) & 641 ICE_RX_FLEX_DESC_PTYPE_M; 642 643 ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype); 644 ice_receive_skb(rx_ring, skb, vlan_tag); 645 } 646 647 ice_finalize_xdp_rx(rx_ring, xdp_xmit); 648 ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes); 649 650 if (xsk_umem_uses_need_wakeup(rx_ring->xsk_umem)) { 651 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use) 652 xsk_set_rx_need_wakeup(rx_ring->xsk_umem); 653 else 654 xsk_clear_rx_need_wakeup(rx_ring->xsk_umem); 655 656 return (int)total_rx_packets; 657 } 658 659 return failure ? budget : (int)total_rx_packets; 660 } 661 662 /** 663 * ice_xmit_zc - Completes AF_XDP entries, and cleans XDP entries 664 * @xdp_ring: XDP Tx ring 665 * @budget: max number of frames to xmit 666 * 667 * Returns true if cleanup/transmission is done. 668 */ 669 static bool ice_xmit_zc(struct ice_ring *xdp_ring, int budget) 670 { 671 struct ice_tx_desc *tx_desc = NULL; 672 bool work_done = true; 673 struct xdp_desc desc; 674 dma_addr_t dma; 675 676 while (likely(budget-- > 0)) { 677 struct ice_tx_buf *tx_buf; 678 679 if (unlikely(!ICE_DESC_UNUSED(xdp_ring))) { 680 xdp_ring->tx_stats.tx_busy++; 681 work_done = false; 682 break; 683 } 684 685 tx_buf = &xdp_ring->tx_buf[xdp_ring->next_to_use]; 686 687 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc)) 688 break; 689 690 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_umem, desc.addr); 691 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_umem, dma, 692 desc.len); 693 694 tx_buf->bytecount = desc.len; 695 696 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use); 697 tx_desc->buf_addr = cpu_to_le64(dma); 698 tx_desc->cmd_type_offset_bsz = 699 ice_build_ctob(ICE_TXD_LAST_DESC_CMD, 0, desc.len, 0); 700 701 xdp_ring->next_to_use++; 702 if (xdp_ring->next_to_use == xdp_ring->count) 703 xdp_ring->next_to_use = 0; 704 } 705 706 if (tx_desc) { 707 ice_xdp_ring_update_tail(xdp_ring); 708 xsk_umem_consume_tx_done(xdp_ring->xsk_umem); 709 if (xsk_umem_uses_need_wakeup(xdp_ring->xsk_umem)) 710 xsk_clear_tx_need_wakeup(xdp_ring->xsk_umem); 711 } 712 713 return budget > 0 && work_done; 714 } 715 716 /** 717 * ice_clean_xdp_tx_buf - Free and unmap XDP Tx buffer 718 * @xdp_ring: XDP Tx ring 719 * @tx_buf: Tx buffer to clean 720 */ 721 static void 722 ice_clean_xdp_tx_buf(struct ice_ring *xdp_ring, struct ice_tx_buf *tx_buf) 723 { 724 xdp_return_frame((struct xdp_frame *)tx_buf->raw_buf); 725 dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma), 726 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE); 727 dma_unmap_len_set(tx_buf, len, 0); 728 } 729 730 /** 731 * ice_clean_tx_irq_zc - Completes AF_XDP entries, and cleans XDP entries 732 * @xdp_ring: XDP Tx ring 733 * @budget: NAPI budget 734 * 735 * Returns true if cleanup/tranmission is done. 736 */ 737 bool ice_clean_tx_irq_zc(struct ice_ring *xdp_ring, int budget) 738 { 739 int total_packets = 0, total_bytes = 0; 740 s16 ntc = xdp_ring->next_to_clean; 741 struct ice_tx_desc *tx_desc; 742 struct ice_tx_buf *tx_buf; 743 u32 xsk_frames = 0; 744 bool xmit_done; 745 746 tx_desc = ICE_TX_DESC(xdp_ring, ntc); 747 tx_buf = &xdp_ring->tx_buf[ntc]; 748 ntc -= xdp_ring->count; 749 750 do { 751 if (!(tx_desc->cmd_type_offset_bsz & 752 cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE))) 753 break; 754 755 total_bytes += tx_buf->bytecount; 756 total_packets++; 757 758 if (tx_buf->raw_buf) { 759 ice_clean_xdp_tx_buf(xdp_ring, tx_buf); 760 tx_buf->raw_buf = NULL; 761 } else { 762 xsk_frames++; 763 } 764 765 tx_desc->cmd_type_offset_bsz = 0; 766 tx_buf++; 767 tx_desc++; 768 ntc++; 769 770 if (unlikely(!ntc)) { 771 ntc -= xdp_ring->count; 772 tx_buf = xdp_ring->tx_buf; 773 tx_desc = ICE_TX_DESC(xdp_ring, 0); 774 } 775 776 prefetch(tx_desc); 777 778 } while (likely(--budget)); 779 780 ntc += xdp_ring->count; 781 xdp_ring->next_to_clean = ntc; 782 783 if (xsk_frames) 784 xsk_umem_complete_tx(xdp_ring->xsk_umem, xsk_frames); 785 786 if (xsk_umem_uses_need_wakeup(xdp_ring->xsk_umem)) { 787 if (xdp_ring->next_to_clean == xdp_ring->next_to_use) 788 xsk_set_tx_need_wakeup(xdp_ring->xsk_umem); 789 else 790 xsk_clear_tx_need_wakeup(xdp_ring->xsk_umem); 791 } 792 793 ice_update_tx_ring_stats(xdp_ring, total_packets, total_bytes); 794 xmit_done = ice_xmit_zc(xdp_ring, ICE_DFLT_IRQ_WORK); 795 796 return budget > 0 && xmit_done; 797 } 798 799 /** 800 * ice_xsk_wakeup - Implements ndo_xsk_wakeup 801 * @netdev: net_device 802 * @queue_id: queue to wake up 803 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI 804 * 805 * Returns negative on error, zero otherwise. 806 */ 807 int 808 ice_xsk_wakeup(struct net_device *netdev, u32 queue_id, 809 u32 __always_unused flags) 810 { 811 struct ice_netdev_priv *np = netdev_priv(netdev); 812 struct ice_q_vector *q_vector; 813 struct ice_vsi *vsi = np->vsi; 814 struct ice_ring *ring; 815 816 if (test_bit(__ICE_DOWN, vsi->state)) 817 return -ENETDOWN; 818 819 if (!ice_is_xdp_ena_vsi(vsi)) 820 return -ENXIO; 821 822 if (queue_id >= vsi->num_txq) 823 return -ENXIO; 824 825 if (!vsi->xdp_rings[queue_id]->xsk_umem) 826 return -ENXIO; 827 828 ring = vsi->xdp_rings[queue_id]; 829 830 /* The idea here is that if NAPI is running, mark a miss, so 831 * it will run again. If not, trigger an interrupt and 832 * schedule the NAPI from interrupt context. If NAPI would be 833 * scheduled here, the interrupt affinity would not be 834 * honored. 835 */ 836 q_vector = ring->q_vector; 837 if (!napi_if_scheduled_mark_missed(&q_vector->napi)) 838 ice_trigger_sw_intr(&vsi->back->hw, q_vector); 839 840 return 0; 841 } 842 843 /** 844 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP UMEM attached 845 * @vsi: VSI to be checked 846 * 847 * Returns true if any of the Rx rings has an AF_XDP UMEM attached 848 */ 849 bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi) 850 { 851 int i; 852 853 if (!vsi->xsk_umems) 854 return false; 855 856 for (i = 0; i < vsi->num_xsk_umems; i++) { 857 if (vsi->xsk_umems[i]) 858 return true; 859 } 860 861 return false; 862 } 863 864 /** 865 * ice_xsk_clean_rx_ring - clean UMEM queues connected to a given Rx ring 866 * @rx_ring: ring to be cleaned 867 */ 868 void ice_xsk_clean_rx_ring(struct ice_ring *rx_ring) 869 { 870 u16 i; 871 872 for (i = 0; i < rx_ring->count; i++) { 873 struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i]; 874 875 if (!rx_buf->xdp) 876 continue; 877 878 rx_buf->xdp = NULL; 879 } 880 } 881 882 /** 883 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its UMEM queues 884 * @xdp_ring: XDP_Tx ring 885 */ 886 void ice_xsk_clean_xdp_ring(struct ice_ring *xdp_ring) 887 { 888 u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use; 889 u32 xsk_frames = 0; 890 891 while (ntc != ntu) { 892 struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc]; 893 894 if (tx_buf->raw_buf) 895 ice_clean_xdp_tx_buf(xdp_ring, tx_buf); 896 else 897 xsk_frames++; 898 899 tx_buf->raw_buf = NULL; 900 901 ntc++; 902 if (ntc >= xdp_ring->count) 903 ntc = 0; 904 } 905 906 if (xsk_frames) 907 xsk_umem_complete_tx(xdp_ring->xsk_umem, xsk_frames); 908 } 909