1 /* Broadcom NetXtreme-C/E network driver. 2 * 3 * Copyright (c) 2016-2017 Broadcom Limited 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License as published by 7 * the Free Software Foundation. 8 */ 9 #include <linux/kernel.h> 10 #include <linux/errno.h> 11 #include <linux/pci.h> 12 #include <linux/netdevice.h> 13 #include <linux/etherdevice.h> 14 #include <linux/if_vlan.h> 15 #include <linux/bpf.h> 16 #include <linux/bpf_trace.h> 17 #include <linux/filter.h> 18 #include <net/page_pool/helpers.h> 19 #include "bnxt_hsi.h" 20 #include "bnxt.h" 21 #include "bnxt_xdp.h" 22 23 DEFINE_STATIC_KEY_FALSE(bnxt_xdp_locking_key); 24 25 struct bnxt_sw_tx_bd *bnxt_xmit_bd(struct bnxt *bp, 26 struct bnxt_tx_ring_info *txr, 27 dma_addr_t mapping, u32 len, 28 struct xdp_buff *xdp) 29 { 30 struct skb_shared_info *sinfo; 31 struct bnxt_sw_tx_bd *tx_buf; 32 struct tx_bd *txbd; 33 int num_frags = 0; 34 u32 flags; 35 u16 prod; 36 int i; 37 38 if (xdp && xdp_buff_has_frags(xdp)) { 39 sinfo = xdp_get_shared_info_from_buff(xdp); 40 num_frags = sinfo->nr_frags; 41 } 42 43 /* fill up the first buffer */ 44 prod = txr->tx_prod; 45 tx_buf = &txr->tx_buf_ring[prod]; 46 tx_buf->nr_frags = num_frags; 47 if (xdp) 48 tx_buf->page = virt_to_head_page(xdp->data); 49 50 txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)]; 51 flags = (len << TX_BD_LEN_SHIFT) | 52 ((num_frags + 1) << TX_BD_FLAGS_BD_CNT_SHIFT) | 53 bnxt_lhint_arr[len >> 9]; 54 txbd->tx_bd_len_flags_type = cpu_to_le32(flags); 55 txbd->tx_bd_opaque = prod; 56 txbd->tx_bd_haddr = cpu_to_le64(mapping); 57 58 /* now let us fill up the frags into the next buffers */ 59 for (i = 0; i < num_frags ; i++) { 60 skb_frag_t *frag = &sinfo->frags[i]; 61 struct bnxt_sw_tx_bd *frag_tx_buf; 62 struct pci_dev *pdev = bp->pdev; 63 dma_addr_t frag_mapping; 64 int frag_len; 65 66 prod = NEXT_TX(prod); 67 WRITE_ONCE(txr->tx_prod, prod); 68 69 /* first fill up the first buffer */ 70 frag_tx_buf = &txr->tx_buf_ring[prod]; 71 frag_tx_buf->page = skb_frag_page(frag); 72 73 txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)]; 74 75 frag_len = skb_frag_size(frag); 76 frag_mapping = skb_frag_dma_map(&pdev->dev, frag, 0, 77 frag_len, DMA_TO_DEVICE); 78 79 if (unlikely(dma_mapping_error(&pdev->dev, frag_mapping))) 80 return NULL; 81 82 dma_unmap_addr_set(frag_tx_buf, mapping, frag_mapping); 83 84 flags = frag_len << TX_BD_LEN_SHIFT; 85 txbd->tx_bd_len_flags_type = cpu_to_le32(flags); 86 txbd->tx_bd_haddr = cpu_to_le64(frag_mapping); 87 88 len = frag_len; 89 } 90 91 flags &= ~TX_BD_LEN; 92 txbd->tx_bd_len_flags_type = cpu_to_le32(((len) << TX_BD_LEN_SHIFT) | flags | 93 TX_BD_FLAGS_PACKET_END); 94 /* Sync TX BD */ 95 wmb(); 96 prod = NEXT_TX(prod); 97 WRITE_ONCE(txr->tx_prod, prod); 98 99 return tx_buf; 100 } 101 102 static void __bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_tx_ring_info *txr, 103 dma_addr_t mapping, u32 len, u16 rx_prod, 104 struct xdp_buff *xdp) 105 { 106 struct bnxt_sw_tx_bd *tx_buf; 107 108 tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, xdp); 109 tx_buf->rx_prod = rx_prod; 110 tx_buf->action = XDP_TX; 111 112 } 113 114 static void __bnxt_xmit_xdp_redirect(struct bnxt *bp, 115 struct bnxt_tx_ring_info *txr, 116 dma_addr_t mapping, u32 len, 117 struct xdp_frame *xdpf) 118 { 119 struct bnxt_sw_tx_bd *tx_buf; 120 121 tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, NULL); 122 tx_buf->action = XDP_REDIRECT; 123 tx_buf->xdpf = xdpf; 124 dma_unmap_addr_set(tx_buf, mapping, mapping); 125 dma_unmap_len_set(tx_buf, len, 0); 126 } 127 128 void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int budget) 129 { 130 struct bnxt_tx_ring_info *txr = bnapi->tx_ring; 131 struct bnxt_rx_ring_info *rxr = bnapi->rx_ring; 132 bool rx_doorbell_needed = false; 133 int nr_pkts = bnapi->tx_pkts; 134 struct bnxt_sw_tx_bd *tx_buf; 135 u16 tx_cons = txr->tx_cons; 136 u16 last_tx_cons = tx_cons; 137 int i, j, frags; 138 139 if (!budget) 140 return; 141 142 for (i = 0; i < nr_pkts; i++) { 143 tx_buf = &txr->tx_buf_ring[tx_cons]; 144 145 if (tx_buf->action == XDP_REDIRECT) { 146 struct pci_dev *pdev = bp->pdev; 147 148 dma_unmap_single(&pdev->dev, 149 dma_unmap_addr(tx_buf, mapping), 150 dma_unmap_len(tx_buf, len), 151 DMA_TO_DEVICE); 152 xdp_return_frame(tx_buf->xdpf); 153 tx_buf->action = 0; 154 tx_buf->xdpf = NULL; 155 } else if (tx_buf->action == XDP_TX) { 156 tx_buf->action = 0; 157 rx_doorbell_needed = true; 158 last_tx_cons = tx_cons; 159 160 frags = tx_buf->nr_frags; 161 for (j = 0; j < frags; j++) { 162 tx_cons = NEXT_TX(tx_cons); 163 tx_buf = &txr->tx_buf_ring[tx_cons]; 164 page_pool_recycle_direct(rxr->page_pool, tx_buf->page); 165 } 166 } else { 167 bnxt_sched_reset_txr(bp, txr, i); 168 return; 169 } 170 tx_cons = NEXT_TX(tx_cons); 171 } 172 173 bnapi->tx_pkts = 0; 174 WRITE_ONCE(txr->tx_cons, tx_cons); 175 if (rx_doorbell_needed) { 176 tx_buf = &txr->tx_buf_ring[last_tx_cons]; 177 bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod); 178 179 } 180 } 181 182 bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr) 183 { 184 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 185 186 return !!xdp_prog; 187 } 188 189 void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, 190 u16 cons, u8 *data_ptr, unsigned int len, 191 struct xdp_buff *xdp) 192 { 193 u32 buflen = BNXT_RX_PAGE_SIZE; 194 struct bnxt_sw_rx_bd *rx_buf; 195 struct pci_dev *pdev; 196 dma_addr_t mapping; 197 u32 offset; 198 199 pdev = bp->pdev; 200 rx_buf = &rxr->rx_buf_ring[cons]; 201 offset = bp->rx_offset; 202 203 mapping = rx_buf->mapping - bp->rx_dma_offset; 204 dma_sync_single_for_cpu(&pdev->dev, mapping + offset, len, bp->rx_dir); 205 206 xdp_init_buff(xdp, buflen, &rxr->xdp_rxq); 207 xdp_prepare_buff(xdp, data_ptr - offset, offset, len, false); 208 } 209 210 void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr, 211 struct xdp_buff *xdp) 212 { 213 struct skb_shared_info *shinfo; 214 int i; 215 216 if (!xdp || !xdp_buff_has_frags(xdp)) 217 return; 218 shinfo = xdp_get_shared_info_from_buff(xdp); 219 for (i = 0; i < shinfo->nr_frags; i++) { 220 struct page *page = skb_frag_page(&shinfo->frags[i]); 221 222 page_pool_recycle_direct(rxr->page_pool, page); 223 } 224 shinfo->nr_frags = 0; 225 } 226 227 /* returns the following: 228 * true - packet consumed by XDP and new buffer is allocated. 229 * false - packet should be passed to the stack. 230 */ 231 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons, 232 struct xdp_buff xdp, struct page *page, u8 **data_ptr, 233 unsigned int *len, u8 *event) 234 { 235 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 236 struct bnxt_tx_ring_info *txr; 237 struct bnxt_sw_rx_bd *rx_buf; 238 struct pci_dev *pdev; 239 dma_addr_t mapping; 240 u32 tx_needed = 1; 241 void *orig_data; 242 u32 tx_avail; 243 u32 offset; 244 u32 act; 245 246 if (!xdp_prog) 247 return false; 248 249 pdev = bp->pdev; 250 offset = bp->rx_offset; 251 252 txr = rxr->bnapi->tx_ring; 253 /* BNXT_RX_PAGE_MODE(bp) when XDP enabled */ 254 orig_data = xdp.data; 255 256 act = bpf_prog_run_xdp(xdp_prog, &xdp); 257 258 tx_avail = bnxt_tx_avail(bp, txr); 259 /* If the tx ring is not full, we must not update the rx producer yet 260 * because we may still be transmitting on some BDs. 261 */ 262 if (tx_avail != bp->tx_ring_size) 263 *event &= ~BNXT_RX_EVENT; 264 265 *len = xdp.data_end - xdp.data; 266 if (orig_data != xdp.data) { 267 offset = xdp.data - xdp.data_hard_start; 268 *data_ptr = xdp.data_hard_start + offset; 269 } 270 271 switch (act) { 272 case XDP_PASS: 273 return false; 274 275 case XDP_TX: 276 rx_buf = &rxr->rx_buf_ring[cons]; 277 mapping = rx_buf->mapping - bp->rx_dma_offset; 278 *event = 0; 279 280 if (unlikely(xdp_buff_has_frags(&xdp))) { 281 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(&xdp); 282 283 tx_needed += sinfo->nr_frags; 284 *event = BNXT_AGG_EVENT; 285 } 286 287 if (tx_avail < tx_needed) { 288 trace_xdp_exception(bp->dev, xdp_prog, act); 289 bnxt_xdp_buff_frags_free(rxr, &xdp); 290 bnxt_reuse_rx_data(rxr, cons, page); 291 return true; 292 } 293 294 dma_sync_single_for_device(&pdev->dev, mapping + offset, *len, 295 bp->rx_dir); 296 297 *event |= BNXT_TX_EVENT; 298 __bnxt_xmit_xdp(bp, txr, mapping + offset, *len, 299 NEXT_RX(rxr->rx_prod), &xdp); 300 bnxt_reuse_rx_data(rxr, cons, page); 301 return true; 302 case XDP_REDIRECT: 303 /* if we are calling this here then we know that the 304 * redirect is coming from a frame received by the 305 * bnxt_en driver. 306 */ 307 rx_buf = &rxr->rx_buf_ring[cons]; 308 mapping = rx_buf->mapping - bp->rx_dma_offset; 309 dma_unmap_page_attrs(&pdev->dev, mapping, 310 BNXT_RX_PAGE_SIZE, bp->rx_dir, 311 DMA_ATTR_WEAK_ORDERING); 312 313 /* if we are unable to allocate a new buffer, abort and reuse */ 314 if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) { 315 trace_xdp_exception(bp->dev, xdp_prog, act); 316 bnxt_xdp_buff_frags_free(rxr, &xdp); 317 bnxt_reuse_rx_data(rxr, cons, page); 318 return true; 319 } 320 321 if (xdp_do_redirect(bp->dev, &xdp, xdp_prog)) { 322 trace_xdp_exception(bp->dev, xdp_prog, act); 323 page_pool_recycle_direct(rxr->page_pool, page); 324 return true; 325 } 326 327 *event |= BNXT_REDIRECT_EVENT; 328 break; 329 default: 330 bpf_warn_invalid_xdp_action(bp->dev, xdp_prog, act); 331 fallthrough; 332 case XDP_ABORTED: 333 trace_xdp_exception(bp->dev, xdp_prog, act); 334 fallthrough; 335 case XDP_DROP: 336 bnxt_xdp_buff_frags_free(rxr, &xdp); 337 bnxt_reuse_rx_data(rxr, cons, page); 338 break; 339 } 340 return true; 341 } 342 343 int bnxt_xdp_xmit(struct net_device *dev, int num_frames, 344 struct xdp_frame **frames, u32 flags) 345 { 346 struct bnxt *bp = netdev_priv(dev); 347 struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog); 348 struct pci_dev *pdev = bp->pdev; 349 struct bnxt_tx_ring_info *txr; 350 dma_addr_t mapping; 351 int nxmit = 0; 352 int ring; 353 int i; 354 355 if (!test_bit(BNXT_STATE_OPEN, &bp->state) || 356 !bp->tx_nr_rings_xdp || 357 !xdp_prog) 358 return -EINVAL; 359 360 ring = smp_processor_id() % bp->tx_nr_rings_xdp; 361 txr = &bp->tx_ring[ring]; 362 363 if (READ_ONCE(txr->dev_state) == BNXT_DEV_STATE_CLOSING) 364 return -EINVAL; 365 366 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 367 spin_lock(&txr->xdp_tx_lock); 368 369 for (i = 0; i < num_frames; i++) { 370 struct xdp_frame *xdp = frames[i]; 371 372 if (!bnxt_tx_avail(bp, txr)) 373 break; 374 375 mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len, 376 DMA_TO_DEVICE); 377 378 if (dma_mapping_error(&pdev->dev, mapping)) 379 break; 380 381 __bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp); 382 nxmit++; 383 } 384 385 if (flags & XDP_XMIT_FLUSH) { 386 /* Sync BD data before updating doorbell */ 387 wmb(); 388 bnxt_db_write(bp, &txr->tx_db, txr->tx_prod); 389 } 390 391 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 392 spin_unlock(&txr->xdp_tx_lock); 393 394 return nxmit; 395 } 396 397 /* Under rtnl_lock */ 398 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog) 399 { 400 struct net_device *dev = bp->dev; 401 int tx_xdp = 0, rc, tc; 402 struct bpf_prog *old; 403 404 if (prog && !prog->aux->xdp_has_frags && 405 bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) { 406 netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n", 407 bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU); 408 return -EOPNOTSUPP; 409 } 410 if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) { 411 netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n"); 412 return -EOPNOTSUPP; 413 } 414 if (prog) 415 tx_xdp = bp->rx_nr_rings; 416 417 tc = netdev_get_num_tc(dev); 418 if (!tc) 419 tc = 1; 420 rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings, 421 true, tc, tx_xdp); 422 if (rc) { 423 netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n"); 424 return rc; 425 } 426 if (netif_running(dev)) 427 bnxt_close_nic(bp, true, false); 428 429 old = xchg(&bp->xdp_prog, prog); 430 if (old) 431 bpf_prog_put(old); 432 433 if (prog) { 434 bnxt_set_rx_skb_mode(bp, true); 435 xdp_features_set_redirect_target(dev, true); 436 } else { 437 int rx, tx; 438 439 xdp_features_clear_redirect_target(dev); 440 bnxt_set_rx_skb_mode(bp, false); 441 bnxt_get_max_rings(bp, &rx, &tx, true); 442 if (rx > 1) { 443 bp->flags &= ~BNXT_FLAG_NO_AGG_RINGS; 444 bp->dev->hw_features |= NETIF_F_LRO; 445 } 446 } 447 bp->tx_nr_rings_xdp = tx_xdp; 448 bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp; 449 bp->cp_nr_rings = max_t(int, bp->tx_nr_rings, bp->rx_nr_rings); 450 bnxt_set_tpa_flags(bp); 451 bnxt_set_ring_params(bp); 452 453 if (netif_running(dev)) 454 return bnxt_open_nic(bp, true, false); 455 456 return 0; 457 } 458 459 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp) 460 { 461 struct bnxt *bp = netdev_priv(dev); 462 int rc; 463 464 switch (xdp->command) { 465 case XDP_SETUP_PROG: 466 rc = bnxt_xdp_set(bp, xdp->prog); 467 break; 468 default: 469 rc = -EINVAL; 470 break; 471 } 472 return rc; 473 } 474 475 struct sk_buff * 476 bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags, 477 struct page_pool *pool, struct xdp_buff *xdp, 478 struct rx_cmp_ext *rxcmp1) 479 { 480 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 481 482 if (!skb) 483 return NULL; 484 skb_checksum_none_assert(skb); 485 if (RX_CMP_L4_CS_OK(rxcmp1)) { 486 if (bp->dev->features & NETIF_F_RXCSUM) { 487 skb->ip_summed = CHECKSUM_UNNECESSARY; 488 skb->csum_level = RX_CMP_ENCAP(rxcmp1); 489 } 490 } 491 xdp_update_skb_shared_info(skb, num_frags, 492 sinfo->xdp_frags_size, 493 BNXT_RX_PAGE_SIZE * sinfo->nr_frags, 494 xdp_buff_is_frag_pfmemalloc(xdp)); 495 return skb; 496 } 497