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 xsk_buff_pool *ixgbe_xsk_pool(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 xsk_get_pool_from_qid(adapter->netdev, qid); 21 } 22 23 static int ixgbe_xsk_pool_enable(struct ixgbe_adapter *adapter, 24 struct xsk_buff_pool *pool, 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_pool_dma_map(pool, &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 clear_bit(qid, adapter->af_xdp_zc_qps); 57 xsk_pool_dma_unmap(pool, IXGBE_RX_DMA_ATTR); 58 return err; 59 } 60 } 61 62 return 0; 63 } 64 65 static int ixgbe_xsk_pool_disable(struct ixgbe_adapter *adapter, u16 qid) 66 { 67 struct xsk_buff_pool *pool; 68 bool if_running; 69 70 pool = xsk_get_pool_from_qid(adapter->netdev, qid); 71 if (!pool) 72 return -EINVAL; 73 74 if_running = netif_running(adapter->netdev) && 75 ixgbe_enabled_xdp_adapter(adapter); 76 77 if (if_running) 78 ixgbe_txrx_ring_disable(adapter, qid); 79 80 clear_bit(qid, adapter->af_xdp_zc_qps); 81 xsk_pool_dma_unmap(pool, IXGBE_RX_DMA_ATTR); 82 83 if (if_running) 84 ixgbe_txrx_ring_enable(adapter, qid); 85 86 return 0; 87 } 88 89 int ixgbe_xsk_pool_setup(struct ixgbe_adapter *adapter, 90 struct xsk_buff_pool *pool, 91 u16 qid) 92 { 93 return pool ? ixgbe_xsk_pool_enable(adapter, pool, qid) : 94 ixgbe_xsk_pool_disable(adapter, qid); 95 } 96 97 static int ixgbe_run_xdp_zc(struct ixgbe_adapter *adapter, 98 struct ixgbe_ring *rx_ring, 99 struct xdp_buff *xdp) 100 { 101 int err, result = IXGBE_XDP_PASS; 102 struct bpf_prog *xdp_prog; 103 struct ixgbe_ring *ring; 104 struct xdp_frame *xdpf; 105 u32 act; 106 107 xdp_prog = READ_ONCE(rx_ring->xdp_prog); 108 act = bpf_prog_run_xdp(xdp_prog, xdp); 109 110 if (likely(act == XDP_REDIRECT)) { 111 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog); 112 if (err) 113 goto out_failure; 114 return IXGBE_XDP_REDIR; 115 } 116 117 switch (act) { 118 case XDP_PASS: 119 break; 120 case XDP_TX: 121 xdpf = xdp_convert_buff_to_frame(xdp); 122 if (unlikely(!xdpf)) 123 goto out_failure; 124 ring = ixgbe_determine_xdp_ring(adapter); 125 if (static_branch_unlikely(&ixgbe_xdp_locking_key)) 126 spin_lock(&ring->tx_lock); 127 result = ixgbe_xmit_xdp_ring(ring, xdpf); 128 if (static_branch_unlikely(&ixgbe_xdp_locking_key)) 129 spin_unlock(&ring->tx_lock); 130 if (result == IXGBE_XDP_CONSUMED) 131 goto out_failure; 132 break; 133 default: 134 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act); 135 fallthrough; 136 case XDP_ABORTED: 137 out_failure: 138 trace_xdp_exception(rx_ring->netdev, xdp_prog, act); 139 fallthrough; /* handle aborts by dropping packet */ 140 case XDP_DROP: 141 result = IXGBE_XDP_CONSUMED; 142 break; 143 } 144 return result; 145 } 146 147 bool ixgbe_alloc_rx_buffers_zc(struct ixgbe_ring *rx_ring, u16 count) 148 { 149 union ixgbe_adv_rx_desc *rx_desc; 150 struct ixgbe_rx_buffer *bi; 151 u16 i = rx_ring->next_to_use; 152 dma_addr_t dma; 153 bool ok = true; 154 155 /* nothing to do */ 156 if (!count) 157 return true; 158 159 rx_desc = IXGBE_RX_DESC(rx_ring, i); 160 bi = &rx_ring->rx_buffer_info[i]; 161 i -= rx_ring->count; 162 163 do { 164 bi->xdp = xsk_buff_alloc(rx_ring->xsk_pool); 165 if (!bi->xdp) { 166 ok = false; 167 break; 168 } 169 170 dma = xsk_buff_xdp_get_dma(bi->xdp); 171 172 /* Refresh the desc even if buffer_addrs didn't change 173 * because each write-back erases this info. 174 */ 175 rx_desc->read.pkt_addr = cpu_to_le64(dma); 176 177 rx_desc++; 178 bi++; 179 i++; 180 if (unlikely(!i)) { 181 rx_desc = IXGBE_RX_DESC(rx_ring, 0); 182 bi = rx_ring->rx_buffer_info; 183 i -= rx_ring->count; 184 } 185 186 /* clear the length for the next_to_use descriptor */ 187 rx_desc->wb.upper.length = 0; 188 189 count--; 190 } while (count); 191 192 i += rx_ring->count; 193 194 if (rx_ring->next_to_use != i) { 195 rx_ring->next_to_use = i; 196 197 /* Force memory writes to complete before letting h/w 198 * know there are new descriptors to fetch. (Only 199 * applicable for weak-ordered memory model archs, 200 * such as IA-64). 201 */ 202 wmb(); 203 writel(i, rx_ring->tail); 204 } 205 206 return ok; 207 } 208 209 static struct sk_buff *ixgbe_construct_skb_zc(struct ixgbe_ring *rx_ring, 210 struct ixgbe_rx_buffer *bi) 211 { 212 unsigned int metasize = bi->xdp->data - bi->xdp->data_meta; 213 unsigned int datasize = bi->xdp->data_end - bi->xdp->data; 214 struct sk_buff *skb; 215 216 /* allocate a skb to store the frags */ 217 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, 218 bi->xdp->data_end - bi->xdp->data_hard_start, 219 GFP_ATOMIC | __GFP_NOWARN); 220 if (unlikely(!skb)) 221 return NULL; 222 223 skb_reserve(skb, bi->xdp->data - bi->xdp->data_hard_start); 224 memcpy(__skb_put(skb, datasize), bi->xdp->data, datasize); 225 if (metasize) 226 skb_metadata_set(skb, metasize); 227 228 xsk_buff_free(bi->xdp); 229 bi->xdp = NULL; 230 return skb; 231 } 232 233 static void ixgbe_inc_ntc(struct ixgbe_ring *rx_ring) 234 { 235 u32 ntc = rx_ring->next_to_clean + 1; 236 237 ntc = (ntc < rx_ring->count) ? ntc : 0; 238 rx_ring->next_to_clean = ntc; 239 prefetch(IXGBE_RX_DESC(rx_ring, ntc)); 240 } 241 242 int ixgbe_clean_rx_irq_zc(struct ixgbe_q_vector *q_vector, 243 struct ixgbe_ring *rx_ring, 244 const int budget) 245 { 246 unsigned int total_rx_bytes = 0, total_rx_packets = 0; 247 struct ixgbe_adapter *adapter = q_vector->adapter; 248 u16 cleaned_count = ixgbe_desc_unused(rx_ring); 249 unsigned int xdp_res, xdp_xmit = 0; 250 bool failure = false; 251 struct sk_buff *skb; 252 253 while (likely(total_rx_packets < budget)) { 254 union ixgbe_adv_rx_desc *rx_desc; 255 struct ixgbe_rx_buffer *bi; 256 unsigned int size; 257 258 /* return some buffers to hardware, one at a time is too slow */ 259 if (cleaned_count >= IXGBE_RX_BUFFER_WRITE) { 260 failure = failure || 261 !ixgbe_alloc_rx_buffers_zc(rx_ring, 262 cleaned_count); 263 cleaned_count = 0; 264 } 265 266 rx_desc = IXGBE_RX_DESC(rx_ring, rx_ring->next_to_clean); 267 size = le16_to_cpu(rx_desc->wb.upper.length); 268 if (!size) 269 break; 270 271 /* This memory barrier is needed to keep us from reading 272 * any other fields out of the rx_desc until we know the 273 * descriptor has been written back 274 */ 275 dma_rmb(); 276 277 bi = &rx_ring->rx_buffer_info[rx_ring->next_to_clean]; 278 279 if (unlikely(!ixgbe_test_staterr(rx_desc, 280 IXGBE_RXD_STAT_EOP))) { 281 struct ixgbe_rx_buffer *next_bi; 282 283 xsk_buff_free(bi->xdp); 284 bi->xdp = NULL; 285 ixgbe_inc_ntc(rx_ring); 286 next_bi = 287 &rx_ring->rx_buffer_info[rx_ring->next_to_clean]; 288 next_bi->discard = true; 289 continue; 290 } 291 292 if (unlikely(bi->discard)) { 293 xsk_buff_free(bi->xdp); 294 bi->xdp = NULL; 295 bi->discard = false; 296 ixgbe_inc_ntc(rx_ring); 297 continue; 298 } 299 300 bi->xdp->data_end = bi->xdp->data + size; 301 xsk_buff_dma_sync_for_cpu(bi->xdp, rx_ring->xsk_pool); 302 xdp_res = ixgbe_run_xdp_zc(adapter, rx_ring, bi->xdp); 303 304 if (xdp_res) { 305 if (xdp_res & (IXGBE_XDP_TX | IXGBE_XDP_REDIR)) 306 xdp_xmit |= xdp_res; 307 else 308 xsk_buff_free(bi->xdp); 309 310 bi->xdp = NULL; 311 total_rx_packets++; 312 total_rx_bytes += size; 313 314 cleaned_count++; 315 ixgbe_inc_ntc(rx_ring); 316 continue; 317 } 318 319 /* XDP_PASS path */ 320 skb = ixgbe_construct_skb_zc(rx_ring, bi); 321 if (!skb) { 322 rx_ring->rx_stats.alloc_rx_buff_failed++; 323 break; 324 } 325 326 cleaned_count++; 327 ixgbe_inc_ntc(rx_ring); 328 329 if (eth_skb_pad(skb)) 330 continue; 331 332 total_rx_bytes += skb->len; 333 total_rx_packets++; 334 335 ixgbe_process_skb_fields(rx_ring, rx_desc, skb); 336 ixgbe_rx_skb(q_vector, skb); 337 } 338 339 if (xdp_xmit & IXGBE_XDP_REDIR) 340 xdp_do_flush_map(); 341 342 if (xdp_xmit & IXGBE_XDP_TX) { 343 struct ixgbe_ring *ring = ixgbe_determine_xdp_ring(adapter); 344 345 ixgbe_xdp_ring_update_tail_locked(ring); 346 } 347 348 u64_stats_update_begin(&rx_ring->syncp); 349 rx_ring->stats.packets += total_rx_packets; 350 rx_ring->stats.bytes += total_rx_bytes; 351 u64_stats_update_end(&rx_ring->syncp); 352 q_vector->rx.total_packets += total_rx_packets; 353 q_vector->rx.total_bytes += total_rx_bytes; 354 355 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) { 356 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use) 357 xsk_set_rx_need_wakeup(rx_ring->xsk_pool); 358 else 359 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool); 360 361 return (int)total_rx_packets; 362 } 363 return failure ? budget : (int)total_rx_packets; 364 } 365 366 void ixgbe_xsk_clean_rx_ring(struct ixgbe_ring *rx_ring) 367 { 368 struct ixgbe_rx_buffer *bi; 369 u16 i; 370 371 for (i = 0; i < rx_ring->count; i++) { 372 bi = &rx_ring->rx_buffer_info[i]; 373 374 if (!bi->xdp) 375 continue; 376 377 xsk_buff_free(bi->xdp); 378 bi->xdp = NULL; 379 } 380 } 381 382 static bool ixgbe_xmit_zc(struct ixgbe_ring *xdp_ring, unsigned int budget) 383 { 384 struct xsk_buff_pool *pool = xdp_ring->xsk_pool; 385 union ixgbe_adv_tx_desc *tx_desc = NULL; 386 struct ixgbe_tx_buffer *tx_bi; 387 bool work_done = true; 388 struct xdp_desc desc; 389 dma_addr_t dma; 390 u32 cmd_type; 391 392 while (budget-- > 0) { 393 if (unlikely(!ixgbe_desc_unused(xdp_ring))) { 394 work_done = false; 395 break; 396 } 397 398 if (!netif_carrier_ok(xdp_ring->netdev)) 399 break; 400 401 if (!xsk_tx_peek_desc(pool, &desc)) 402 break; 403 404 dma = xsk_buff_raw_get_dma(pool, desc.addr); 405 xsk_buff_raw_dma_sync_for_device(pool, dma, desc.len); 406 407 tx_bi = &xdp_ring->tx_buffer_info[xdp_ring->next_to_use]; 408 tx_bi->bytecount = desc.len; 409 tx_bi->xdpf = NULL; 410 tx_bi->gso_segs = 1; 411 412 tx_desc = IXGBE_TX_DESC(xdp_ring, xdp_ring->next_to_use); 413 tx_desc->read.buffer_addr = cpu_to_le64(dma); 414 415 /* put descriptor type bits */ 416 cmd_type = IXGBE_ADVTXD_DTYP_DATA | 417 IXGBE_ADVTXD_DCMD_DEXT | 418 IXGBE_ADVTXD_DCMD_IFCS; 419 cmd_type |= desc.len | IXGBE_TXD_CMD; 420 tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type); 421 tx_desc->read.olinfo_status = 422 cpu_to_le32(desc.len << IXGBE_ADVTXD_PAYLEN_SHIFT); 423 424 xdp_ring->next_to_use++; 425 if (xdp_ring->next_to_use == xdp_ring->count) 426 xdp_ring->next_to_use = 0; 427 } 428 429 if (tx_desc) { 430 ixgbe_xdp_ring_update_tail(xdp_ring); 431 xsk_tx_release(pool); 432 } 433 434 return !!budget && work_done; 435 } 436 437 static void ixgbe_clean_xdp_tx_buffer(struct ixgbe_ring *tx_ring, 438 struct ixgbe_tx_buffer *tx_bi) 439 { 440 xdp_return_frame(tx_bi->xdpf); 441 dma_unmap_single(tx_ring->dev, 442 dma_unmap_addr(tx_bi, dma), 443 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE); 444 dma_unmap_len_set(tx_bi, len, 0); 445 } 446 447 bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector, 448 struct ixgbe_ring *tx_ring, int napi_budget) 449 { 450 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 451 unsigned int total_packets = 0, total_bytes = 0; 452 struct xsk_buff_pool *pool = tx_ring->xsk_pool; 453 union ixgbe_adv_tx_desc *tx_desc; 454 struct ixgbe_tx_buffer *tx_bi; 455 u32 xsk_frames = 0; 456 457 tx_bi = &tx_ring->tx_buffer_info[ntc]; 458 tx_desc = IXGBE_TX_DESC(tx_ring, ntc); 459 460 while (ntc != ntu) { 461 if (!(tx_desc->wb.status & cpu_to_le32(IXGBE_TXD_STAT_DD))) 462 break; 463 464 total_bytes += tx_bi->bytecount; 465 total_packets += tx_bi->gso_segs; 466 467 if (tx_bi->xdpf) 468 ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi); 469 else 470 xsk_frames++; 471 472 tx_bi->xdpf = NULL; 473 474 tx_bi++; 475 tx_desc++; 476 ntc++; 477 if (unlikely(ntc == tx_ring->count)) { 478 ntc = 0; 479 tx_bi = tx_ring->tx_buffer_info; 480 tx_desc = IXGBE_TX_DESC(tx_ring, 0); 481 } 482 483 /* issue prefetch for next Tx descriptor */ 484 prefetch(tx_desc); 485 } 486 487 tx_ring->next_to_clean = ntc; 488 489 u64_stats_update_begin(&tx_ring->syncp); 490 tx_ring->stats.bytes += total_bytes; 491 tx_ring->stats.packets += total_packets; 492 u64_stats_update_end(&tx_ring->syncp); 493 q_vector->tx.total_bytes += total_bytes; 494 q_vector->tx.total_packets += total_packets; 495 496 if (xsk_frames) 497 xsk_tx_completed(pool, xsk_frames); 498 499 if (xsk_uses_need_wakeup(pool)) 500 xsk_set_tx_need_wakeup(pool); 501 502 return ixgbe_xmit_zc(tx_ring, q_vector->tx.work_limit); 503 } 504 505 int ixgbe_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags) 506 { 507 struct ixgbe_adapter *adapter = netdev_priv(dev); 508 struct ixgbe_ring *ring; 509 510 if (test_bit(__IXGBE_DOWN, &adapter->state)) 511 return -ENETDOWN; 512 513 if (!READ_ONCE(adapter->xdp_prog)) 514 return -ENXIO; 515 516 if (qid >= adapter->num_xdp_queues) 517 return -ENXIO; 518 519 ring = adapter->xdp_ring[qid]; 520 521 if (test_bit(__IXGBE_TX_DISABLED, &ring->state)) 522 return -ENETDOWN; 523 524 if (!ring->xsk_pool) 525 return -ENXIO; 526 527 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi)) { 528 u64 eics = BIT_ULL(ring->q_vector->v_idx); 529 530 ixgbe_irq_rearm_queues(adapter, eics); 531 } 532 533 return 0; 534 } 535 536 void ixgbe_xsk_clean_tx_ring(struct ixgbe_ring *tx_ring) 537 { 538 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use; 539 struct xsk_buff_pool *pool = tx_ring->xsk_pool; 540 struct ixgbe_tx_buffer *tx_bi; 541 u32 xsk_frames = 0; 542 543 while (ntc != ntu) { 544 tx_bi = &tx_ring->tx_buffer_info[ntc]; 545 546 if (tx_bi->xdpf) 547 ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi); 548 else 549 xsk_frames++; 550 551 tx_bi->xdpf = NULL; 552 553 ntc++; 554 if (ntc == tx_ring->count) 555 ntc = 0; 556 } 557 558 if (xsk_frames) 559 xsk_tx_completed(pool, xsk_frames); 560 } 561