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 "ixgbe.h" 9 #include "ixgbe_txrx_common.h" 10 11 struct xdp_umem *ixgbe_xsk_umem(struct ixgbe_adapter *adapter, 12 struct ixgbe_ring *ring) 13 { 14 bool xdp_on = READ_ONCE(adapter->xdp_prog); 15 int qid = ring->ring_idx; 16 17 if (!xdp_on || !test_bit(qid, adapter->af_xdp_zc_qps)) 18 return NULL; 19 20 return xdp_get_umem_from_qid(adapter->netdev, qid); 21 } 22 23 static int ixgbe_xsk_umem_enable(struct ixgbe_adapter *adapter, 24 struct xdp_umem *umem, 25 u16 qid) 26 { 27 struct net_device *netdev = adapter->netdev; 28 bool if_running; 29 int err; 30 31 if (qid >= adapter->num_rx_queues) 32 return -EINVAL; 33 34 if (qid >= netdev->real_num_rx_queues || 35 qid >= netdev->real_num_tx_queues) 36 return -EINVAL; 37 38 err = xsk_buff_dma_map(umem, &adapter->pdev->dev, IXGBE_RX_DMA_ATTR); 39 if (err) 40 return err; 41 42 if_running = netif_running(adapter->netdev) && 43 ixgbe_enabled_xdp_adapter(adapter); 44 45 if (if_running) 46 ixgbe_txrx_ring_disable(adapter, qid); 47 48 set_bit(qid, adapter->af_xdp_zc_qps); 49 50 if (if_running) { 51 ixgbe_txrx_ring_enable(adapter, qid); 52 53 /* Kick start the NAPI context so that receiving will start */ 54 err = ixgbe_xsk_wakeup(adapter->netdev, qid, XDP_WAKEUP_RX); 55 if (err) 56 return err; 57 } 58 59 return 0; 60 } 61 62 static int ixgbe_xsk_umem_disable(struct ixgbe_adapter *adapter, u16 qid) 63 { 64 struct xdp_umem *umem; 65 bool if_running; 66 67 umem = xdp_get_umem_from_qid(adapter->netdev, qid); 68 if (!umem) 69 return -EINVAL; 70 71 if_running = netif_running(adapter->netdev) && 72 ixgbe_enabled_xdp_adapter(adapter); 73 74 if (if_running) 75 ixgbe_txrx_ring_disable(adapter, qid); 76 77 clear_bit(qid, adapter->af_xdp_zc_qps); 78 xsk_buff_dma_unmap(umem, IXGBE_RX_DMA_ATTR); 79 80 if (if_running) 81 ixgbe_txrx_ring_enable(adapter, qid); 82 83 return 0; 84 } 85 86 int ixgbe_xsk_umem_setup(struct ixgbe_adapter *adapter, struct xdp_umem *umem, 87 u16 qid) 88 { 89 return umem ? ixgbe_xsk_umem_enable(adapter, umem, qid) : 90 ixgbe_xsk_umem_disable(adapter, qid); 91 } 92 93 static int ixgbe_run_xdp_zc(struct ixgbe_adapter *adapter, 94 struct ixgbe_ring *rx_ring, 95 struct xdp_buff *xdp) 96 { 97 int err, result = IXGBE_XDP_PASS; 98 struct bpf_prog *xdp_prog; 99 struct xdp_frame *xdpf; 100 u32 act; 101 102 rcu_read_lock(); 103 xdp_prog = READ_ONCE(rx_ring->xdp_prog); 104 act = bpf_prog_run_xdp(xdp_prog, xdp); 105 106 switch (act) { 107 case XDP_PASS: 108 break; 109 case XDP_TX: 110 xdpf = xdp_convert_buff_to_frame(xdp); 111 if (unlikely(!xdpf)) { 112 result = IXGBE_XDP_CONSUMED; 113 break; 114 } 115 result = ixgbe_xmit_xdp_ring(adapter, xdpf); 116 break; 117 case XDP_REDIRECT: 118 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog); 119 result = !err ? IXGBE_XDP_REDIR : IXGBE_XDP_CONSUMED; 120 break; 121 default: 122 bpf_warn_invalid_xdp_action(act); 123 /* fallthrough */ 124 case XDP_ABORTED: 125 trace_xdp_exception(rx_ring->netdev, xdp_prog, act); 126 /* fallthrough -- handle aborts by dropping packet */ 127 case XDP_DROP: 128 result = IXGBE_XDP_CONSUMED; 129 break; 130 } 131 rcu_read_unlock(); 132 return result; 133 } 134 135 bool ixgbe_alloc_rx_buffers_zc(struct ixgbe_ring *rx_ring, u16 count) 136 { 137 union ixgbe_adv_rx_desc *rx_desc; 138 struct ixgbe_rx_buffer *bi; 139 u16 i = rx_ring->next_to_use; 140 dma_addr_t dma; 141 bool ok = true; 142 143 /* nothing to do */ 144 if (!count) 145 return true; 146 147 rx_desc = IXGBE_RX_DESC(rx_ring, i); 148 bi = &rx_ring->rx_buffer_info[i]; 149 i -= rx_ring->count; 150 151 do { 152 bi->xdp = xsk_buff_alloc(rx_ring->xsk_umem); 153 if (!bi->xdp) { 154 ok = false; 155 break; 156 } 157 158 dma = xsk_buff_xdp_get_dma(bi->xdp); 159 160 /* Refresh the desc even if buffer_addrs didn't change 161 * because each write-back erases this info. 162 */ 163 rx_desc->read.pkt_addr = cpu_to_le64(dma); 164 165 rx_desc++; 166 bi++; 167 i++; 168 if (unlikely(!i)) { 169 rx_desc = IXGBE_RX_DESC(rx_ring, 0); 170 bi = rx_ring->rx_buffer_info; 171 i -= rx_ring->count; 172 } 173 174 /* clear the length for the next_to_use descriptor */ 175 rx_desc->wb.upper.length = 0; 176 177 count--; 178 } while (count); 179 180 i += rx_ring->count; 181 182 if (rx_ring->next_to_use != i) { 183 rx_ring->next_to_use = i; 184 185 /* Force memory writes to complete before letting h/w 186 * know there are new descriptors to fetch. (Only 187 * applicable for weak-ordered memory model archs, 188 * such as IA-64). 189 */ 190 wmb(); 191 writel(i, rx_ring->tail); 192 } 193 194 return ok; 195 } 196 197 static struct sk_buff *ixgbe_construct_skb_zc(struct ixgbe_ring *rx_ring, 198 struct ixgbe_rx_buffer *bi) 199 { 200 unsigned int metasize = bi->xdp->data - bi->xdp->data_meta; 201 unsigned int datasize = bi->xdp->data_end - bi->xdp->data; 202 struct sk_buff *skb; 203 204 /* allocate a skb to store the frags */ 205 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, 206 bi->xdp->data_end - bi->xdp->data_hard_start, 207 GFP_ATOMIC | __GFP_NOWARN); 208 if (unlikely(!skb)) 209 return NULL; 210 211 skb_reserve(skb, bi->xdp->data - bi->xdp->data_hard_start); 212 memcpy(__skb_put(skb, datasize), bi->xdp->data, datasize); 213 if (metasize) 214 skb_metadata_set(skb, metasize); 215 216 xsk_buff_free(bi->xdp); 217 bi->xdp = NULL; 218 return skb; 219 } 220 221 static void ixgbe_inc_ntc(struct ixgbe_ring *rx_ring) 222 { 223 u32 ntc = rx_ring->next_to_clean + 1; 224 225 ntc = (ntc < rx_ring->count) ? ntc : 0; 226 rx_ring->next_to_clean = ntc; 227 prefetch(IXGBE_RX_DESC(rx_ring, ntc)); 228 } 229 230 int ixgbe_clean_rx_irq_zc(struct ixgbe_q_vector *q_vector, 231 struct ixgbe_ring *rx_ring, 232 const int budget) 233 { 234 unsigned int total_rx_bytes = 0, total_rx_packets = 0; 235 struct ixgbe_adapter *adapter = q_vector->adapter; 236 u16 cleaned_count = ixgbe_desc_unused(rx_ring); 237 unsigned int xdp_res, xdp_xmit = 0; 238 bool failure = false; 239 struct sk_buff *skb; 240 241 while (likely(total_rx_packets < budget)) { 242 union ixgbe_adv_rx_desc *rx_desc; 243 struct ixgbe_rx_buffer *bi; 244 unsigned int size; 245 246 /* return some buffers to hardware, one at a time is too slow */ 247 if (cleaned_count >= IXGBE_RX_BUFFER_WRITE) { 248 failure = failure || 249 !ixgbe_alloc_rx_buffers_zc(rx_ring, 250 cleaned_count); 251 cleaned_count = 0; 252 } 253 254 rx_desc = IXGBE_RX_DESC(rx_ring, rx_ring->next_to_clean); 255 size = le16_to_cpu(rx_desc->wb.upper.length); 256 if (!size) 257 break; 258 259 /* This memory barrier is needed to keep us from reading 260 * any other fields out of the rx_desc until we know the 261 * descriptor has been written back 262 */ 263 dma_rmb(); 264 265 bi = &rx_ring->rx_buffer_info[rx_ring->next_to_clean]; 266 267 if (unlikely(!ixgbe_test_staterr(rx_desc, 268 IXGBE_RXD_STAT_EOP))) { 269 struct ixgbe_rx_buffer *next_bi; 270 271 xsk_buff_free(bi->xdp); 272 bi->xdp = NULL; 273 ixgbe_inc_ntc(rx_ring); 274 next_bi = 275 &rx_ring->rx_buffer_info[rx_ring->next_to_clean]; 276 next_bi->discard = true; 277 continue; 278 } 279 280 if (unlikely(bi->discard)) { 281 xsk_buff_free(bi->xdp); 282 bi->xdp = NULL; 283 bi->discard = false; 284 ixgbe_inc_ntc(rx_ring); 285 continue; 286 } 287 288 bi->xdp->data_end = bi->xdp->data + size; 289 xsk_buff_dma_sync_for_cpu(bi->xdp); 290 xdp_res = ixgbe_run_xdp_zc(adapter, rx_ring, bi->xdp); 291 292 if (xdp_res) { 293 if (xdp_res & (IXGBE_XDP_TX | IXGBE_XDP_REDIR)) 294 xdp_xmit |= xdp_res; 295 else 296 xsk_buff_free(bi->xdp); 297 298 bi->xdp = NULL; 299 total_rx_packets++; 300 total_rx_bytes += size; 301 302 cleaned_count++; 303 ixgbe_inc_ntc(rx_ring); 304 continue; 305 } 306 307 /* XDP_PASS path */ 308 skb = ixgbe_construct_skb_zc(rx_ring, bi); 309 if (!skb) { 310 rx_ring->rx_stats.alloc_rx_buff_failed++; 311 break; 312 } 313 314 cleaned_count++; 315 ixgbe_inc_ntc(rx_ring); 316 317 if (eth_skb_pad(skb)) 318 continue; 319 320 total_rx_bytes += skb->len; 321 total_rx_packets++; 322 323 ixgbe_process_skb_fields(rx_ring, rx_desc, skb); 324 ixgbe_rx_skb(q_vector, skb); 325 } 326 327 if (xdp_xmit & IXGBE_XDP_REDIR) 328 xdp_do_flush_map(); 329 330 if (xdp_xmit & IXGBE_XDP_TX) { 331 struct ixgbe_ring *ring = adapter->xdp_ring[smp_processor_id()]; 332 333 /* Force memory writes to complete before letting h/w 334 * know there are new descriptors to fetch. 335 */ 336 wmb(); 337 writel(ring->next_to_use, ring->tail); 338 } 339 340 u64_stats_update_begin(&rx_ring->syncp); 341 rx_ring->stats.packets += total_rx_packets; 342 rx_ring->stats.bytes += total_rx_bytes; 343 u64_stats_update_end(&rx_ring->syncp); 344 q_vector->rx.total_packets += total_rx_packets; 345 q_vector->rx.total_bytes += total_rx_bytes; 346 347 if (xsk_umem_uses_need_wakeup(rx_ring->xsk_umem)) { 348 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use) 349 xsk_set_rx_need_wakeup(rx_ring->xsk_umem); 350 else 351 xsk_clear_rx_need_wakeup(rx_ring->xsk_umem); 352 353 return (int)total_rx_packets; 354 } 355 return failure ? budget : (int)total_rx_packets; 356 } 357 358 void ixgbe_xsk_clean_rx_ring(struct ixgbe_ring *rx_ring) 359 { 360 struct ixgbe_rx_buffer *bi; 361 u16 i; 362 363 for (i = 0; i < rx_ring->count; i++) { 364 bi = &rx_ring->rx_buffer_info[i]; 365 366 if (!bi->xdp) 367 continue; 368 369 xsk_buff_free(bi->xdp); 370 bi->xdp = NULL; 371 } 372 } 373 374 static bool ixgbe_xmit_zc(struct ixgbe_ring *xdp_ring, unsigned int budget) 375 { 376 union ixgbe_adv_tx_desc *tx_desc = NULL; 377 struct ixgbe_tx_buffer *tx_bi; 378 bool work_done = true; 379 struct xdp_desc desc; 380 dma_addr_t dma; 381 u32 cmd_type; 382 383 while (budget-- > 0) { 384 if (unlikely(!ixgbe_desc_unused(xdp_ring)) || 385 !netif_carrier_ok(xdp_ring->netdev)) { 386 work_done = false; 387 break; 388 } 389 390 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc)) 391 break; 392 393 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_umem, desc.addr); 394 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_umem, dma, 395 desc.len); 396 397 tx_bi = &xdp_ring->tx_buffer_info[xdp_ring->next_to_use]; 398 tx_bi->bytecount = desc.len; 399 tx_bi->xdpf = NULL; 400 tx_bi->gso_segs = 1; 401 402 tx_desc = IXGBE_TX_DESC(xdp_ring, xdp_ring->next_to_use); 403 tx_desc->read.buffer_addr = cpu_to_le64(dma); 404 405 /* put descriptor type bits */ 406 cmd_type = IXGBE_ADVTXD_DTYP_DATA | 407 IXGBE_ADVTXD_DCMD_DEXT | 408 IXGBE_ADVTXD_DCMD_IFCS; 409 cmd_type |= desc.len | IXGBE_TXD_CMD; 410 tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type); 411 tx_desc->read.olinfo_status = 412 cpu_to_le32(desc.len << IXGBE_ADVTXD_PAYLEN_SHIFT); 413 414 xdp_ring->next_to_use++; 415 if (xdp_ring->next_to_use == xdp_ring->count) 416 xdp_ring->next_to_use = 0; 417 } 418 419 if (tx_desc) { 420 ixgbe_xdp_ring_update_tail(xdp_ring); 421 xsk_umem_consume_tx_done(xdp_ring->xsk_umem); 422 } 423 424 return !!budget && work_done; 425 } 426 427 static void ixgbe_clean_xdp_tx_buffer(struct ixgbe_ring *tx_ring, 428 struct ixgbe_tx_buffer *tx_bi) 429 { 430 xdp_return_frame(tx_bi->xdpf); 431 dma_unmap_single(tx_ring->dev, 432 dma_unmap_addr(tx_bi, dma), 433 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE); 434 dma_unmap_len_set(tx_bi, len, 0); 435 } 436 437 bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector, 438 struct ixgbe_ring *tx_ring, int napi_budget) 439 { 440 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 441 unsigned int total_packets = 0, total_bytes = 0; 442 struct xdp_umem *umem = tx_ring->xsk_umem; 443 union ixgbe_adv_tx_desc *tx_desc; 444 struct ixgbe_tx_buffer *tx_bi; 445 u32 xsk_frames = 0; 446 447 tx_bi = &tx_ring->tx_buffer_info[ntc]; 448 tx_desc = IXGBE_TX_DESC(tx_ring, ntc); 449 450 while (ntc != ntu) { 451 if (!(tx_desc->wb.status & cpu_to_le32(IXGBE_TXD_STAT_DD))) 452 break; 453 454 total_bytes += tx_bi->bytecount; 455 total_packets += tx_bi->gso_segs; 456 457 if (tx_bi->xdpf) 458 ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi); 459 else 460 xsk_frames++; 461 462 tx_bi->xdpf = NULL; 463 464 tx_bi++; 465 tx_desc++; 466 ntc++; 467 if (unlikely(ntc == tx_ring->count)) { 468 ntc = 0; 469 tx_bi = tx_ring->tx_buffer_info; 470 tx_desc = IXGBE_TX_DESC(tx_ring, 0); 471 } 472 473 /* issue prefetch for next Tx descriptor */ 474 prefetch(tx_desc); 475 } 476 477 tx_ring->next_to_clean = ntc; 478 479 u64_stats_update_begin(&tx_ring->syncp); 480 tx_ring->stats.bytes += total_bytes; 481 tx_ring->stats.packets += total_packets; 482 u64_stats_update_end(&tx_ring->syncp); 483 q_vector->tx.total_bytes += total_bytes; 484 q_vector->tx.total_packets += total_packets; 485 486 if (xsk_frames) 487 xsk_umem_complete_tx(umem, xsk_frames); 488 489 if (xsk_umem_uses_need_wakeup(tx_ring->xsk_umem)) 490 xsk_set_tx_need_wakeup(tx_ring->xsk_umem); 491 492 return ixgbe_xmit_zc(tx_ring, q_vector->tx.work_limit); 493 } 494 495 int ixgbe_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags) 496 { 497 struct ixgbe_adapter *adapter = netdev_priv(dev); 498 struct ixgbe_ring *ring; 499 500 if (test_bit(__IXGBE_DOWN, &adapter->state)) 501 return -ENETDOWN; 502 503 if (!READ_ONCE(adapter->xdp_prog)) 504 return -ENXIO; 505 506 if (qid >= adapter->num_xdp_queues) 507 return -ENXIO; 508 509 ring = adapter->xdp_ring[qid]; 510 511 if (test_bit(__IXGBE_TX_DISABLED, &ring->state)) 512 return -ENETDOWN; 513 514 if (!ring->xsk_umem) 515 return -ENXIO; 516 517 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi)) { 518 u64 eics = BIT_ULL(ring->q_vector->v_idx); 519 520 ixgbe_irq_rearm_queues(adapter, eics); 521 } 522 523 return 0; 524 } 525 526 void ixgbe_xsk_clean_tx_ring(struct ixgbe_ring *tx_ring) 527 { 528 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 529 struct xdp_umem *umem = tx_ring->xsk_umem; 530 struct ixgbe_tx_buffer *tx_bi; 531 u32 xsk_frames = 0; 532 533 while (ntc != ntu) { 534 tx_bi = &tx_ring->tx_buffer_info[ntc]; 535 536 if (tx_bi->xdpf) 537 ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi); 538 else 539 xsk_frames++; 540 541 tx_bi->xdpf = NULL; 542 543 ntc++; 544 if (ntc == tx_ring->count) 545 ntc = 0; 546 } 547 548 if (xsk_frames) 549 xsk_umem_complete_tx(umem, xsk_frames); 550 } 551