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.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 rx_doorbell_needed = true; 157 last_tx_cons = tx_cons; 158 159 frags = tx_buf->nr_frags; 160 for (j = 0; j < frags; j++) { 161 tx_cons = NEXT_TX(tx_cons); 162 tx_buf = &txr->tx_buf_ring[tx_cons]; 163 page_pool_recycle_direct(rxr->page_pool, tx_buf->page); 164 } 165 } 166 tx_cons = NEXT_TX(tx_cons); 167 } 168 169 bnapi->tx_pkts = 0; 170 WRITE_ONCE(txr->tx_cons, tx_cons); 171 if (rx_doorbell_needed) { 172 tx_buf = &txr->tx_buf_ring[last_tx_cons]; 173 bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod); 174 175 } 176 } 177 178 bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr) 179 { 180 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 181 182 return !!xdp_prog; 183 } 184 185 void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, 186 u16 cons, u8 *data_ptr, unsigned int len, 187 struct xdp_buff *xdp) 188 { 189 u32 buflen = BNXT_RX_PAGE_SIZE; 190 struct bnxt_sw_rx_bd *rx_buf; 191 struct pci_dev *pdev; 192 dma_addr_t mapping; 193 u32 offset; 194 195 pdev = bp->pdev; 196 rx_buf = &rxr->rx_buf_ring[cons]; 197 offset = bp->rx_offset; 198 199 mapping = rx_buf->mapping - bp->rx_dma_offset; 200 dma_sync_single_for_cpu(&pdev->dev, mapping + offset, len, bp->rx_dir); 201 202 xdp_init_buff(xdp, buflen, &rxr->xdp_rxq); 203 xdp_prepare_buff(xdp, data_ptr - offset, offset, len, false); 204 } 205 206 void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr, 207 struct xdp_buff *xdp) 208 { 209 struct skb_shared_info *shinfo; 210 int i; 211 212 if (!xdp || !xdp_buff_has_frags(xdp)) 213 return; 214 shinfo = xdp_get_shared_info_from_buff(xdp); 215 for (i = 0; i < shinfo->nr_frags; i++) { 216 struct page *page = skb_frag_page(&shinfo->frags[i]); 217 218 page_pool_recycle_direct(rxr->page_pool, page); 219 } 220 shinfo->nr_frags = 0; 221 } 222 223 /* returns the following: 224 * true - packet consumed by XDP and new buffer is allocated. 225 * false - packet should be passed to the stack. 226 */ 227 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons, 228 struct xdp_buff xdp, struct page *page, u8 **data_ptr, 229 unsigned int *len, u8 *event) 230 { 231 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 232 struct bnxt_tx_ring_info *txr; 233 struct bnxt_sw_rx_bd *rx_buf; 234 struct pci_dev *pdev; 235 dma_addr_t mapping; 236 u32 tx_needed = 1; 237 void *orig_data; 238 u32 tx_avail; 239 u32 offset; 240 u32 act; 241 242 if (!xdp_prog) 243 return false; 244 245 pdev = bp->pdev; 246 offset = bp->rx_offset; 247 248 txr = rxr->bnapi->tx_ring; 249 /* BNXT_RX_PAGE_MODE(bp) when XDP enabled */ 250 orig_data = xdp.data; 251 252 act = bpf_prog_run_xdp(xdp_prog, &xdp); 253 254 tx_avail = bnxt_tx_avail(bp, txr); 255 /* If the tx ring is not full, we must not update the rx producer yet 256 * because we may still be transmitting on some BDs. 257 */ 258 if (tx_avail != bp->tx_ring_size) 259 *event &= ~BNXT_RX_EVENT; 260 261 *len = xdp.data_end - xdp.data; 262 if (orig_data != xdp.data) { 263 offset = xdp.data - xdp.data_hard_start; 264 *data_ptr = xdp.data_hard_start + offset; 265 } 266 267 switch (act) { 268 case XDP_PASS: 269 return false; 270 271 case XDP_TX: 272 rx_buf = &rxr->rx_buf_ring[cons]; 273 mapping = rx_buf->mapping - bp->rx_dma_offset; 274 *event = 0; 275 276 if (unlikely(xdp_buff_has_frags(&xdp))) { 277 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(&xdp); 278 279 tx_needed += sinfo->nr_frags; 280 *event = BNXT_AGG_EVENT; 281 } 282 283 if (tx_avail < tx_needed) { 284 trace_xdp_exception(bp->dev, xdp_prog, act); 285 bnxt_xdp_buff_frags_free(rxr, &xdp); 286 bnxt_reuse_rx_data(rxr, cons, page); 287 return true; 288 } 289 290 dma_sync_single_for_device(&pdev->dev, mapping + offset, *len, 291 bp->rx_dir); 292 293 *event |= BNXT_TX_EVENT; 294 __bnxt_xmit_xdp(bp, txr, mapping + offset, *len, 295 NEXT_RX(rxr->rx_prod), &xdp); 296 bnxt_reuse_rx_data(rxr, cons, page); 297 return true; 298 case XDP_REDIRECT: 299 /* if we are calling this here then we know that the 300 * redirect is coming from a frame received by the 301 * bnxt_en driver. 302 */ 303 rx_buf = &rxr->rx_buf_ring[cons]; 304 mapping = rx_buf->mapping - bp->rx_dma_offset; 305 dma_unmap_page_attrs(&pdev->dev, mapping, 306 BNXT_RX_PAGE_SIZE, bp->rx_dir, 307 DMA_ATTR_WEAK_ORDERING); 308 309 /* if we are unable to allocate a new buffer, abort and reuse */ 310 if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) { 311 trace_xdp_exception(bp->dev, xdp_prog, act); 312 bnxt_xdp_buff_frags_free(rxr, &xdp); 313 bnxt_reuse_rx_data(rxr, cons, page); 314 return true; 315 } 316 317 if (xdp_do_redirect(bp->dev, &xdp, xdp_prog)) { 318 trace_xdp_exception(bp->dev, xdp_prog, act); 319 page_pool_recycle_direct(rxr->page_pool, page); 320 return true; 321 } 322 323 *event |= BNXT_REDIRECT_EVENT; 324 break; 325 default: 326 bpf_warn_invalid_xdp_action(bp->dev, xdp_prog, act); 327 fallthrough; 328 case XDP_ABORTED: 329 trace_xdp_exception(bp->dev, xdp_prog, act); 330 fallthrough; 331 case XDP_DROP: 332 bnxt_xdp_buff_frags_free(rxr, &xdp); 333 bnxt_reuse_rx_data(rxr, cons, page); 334 break; 335 } 336 return true; 337 } 338 339 int bnxt_xdp_xmit(struct net_device *dev, int num_frames, 340 struct xdp_frame **frames, u32 flags) 341 { 342 struct bnxt *bp = netdev_priv(dev); 343 struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog); 344 struct pci_dev *pdev = bp->pdev; 345 struct bnxt_tx_ring_info *txr; 346 dma_addr_t mapping; 347 int nxmit = 0; 348 int ring; 349 int i; 350 351 if (!test_bit(BNXT_STATE_OPEN, &bp->state) || 352 !bp->tx_nr_rings_xdp || 353 !xdp_prog) 354 return -EINVAL; 355 356 ring = smp_processor_id() % bp->tx_nr_rings_xdp; 357 txr = &bp->tx_ring[ring]; 358 359 if (READ_ONCE(txr->dev_state) == BNXT_DEV_STATE_CLOSING) 360 return -EINVAL; 361 362 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 363 spin_lock(&txr->xdp_tx_lock); 364 365 for (i = 0; i < num_frames; i++) { 366 struct xdp_frame *xdp = frames[i]; 367 368 if (!bnxt_tx_avail(bp, txr)) 369 break; 370 371 mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len, 372 DMA_TO_DEVICE); 373 374 if (dma_mapping_error(&pdev->dev, mapping)) 375 break; 376 377 __bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp); 378 nxmit++; 379 } 380 381 if (flags & XDP_XMIT_FLUSH) { 382 /* Sync BD data before updating doorbell */ 383 wmb(); 384 bnxt_db_write(bp, &txr->tx_db, txr->tx_prod); 385 } 386 387 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 388 spin_unlock(&txr->xdp_tx_lock); 389 390 return nxmit; 391 } 392 393 /* Under rtnl_lock */ 394 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog) 395 { 396 struct net_device *dev = bp->dev; 397 int tx_xdp = 0, rc, tc; 398 struct bpf_prog *old; 399 400 if (prog && !prog->aux->xdp_has_frags && 401 bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) { 402 netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n", 403 bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU); 404 return -EOPNOTSUPP; 405 } 406 if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) { 407 netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n"); 408 return -EOPNOTSUPP; 409 } 410 if (prog) 411 tx_xdp = bp->rx_nr_rings; 412 413 tc = netdev_get_num_tc(dev); 414 if (!tc) 415 tc = 1; 416 rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings, 417 true, tc, tx_xdp); 418 if (rc) { 419 netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n"); 420 return rc; 421 } 422 if (netif_running(dev)) 423 bnxt_close_nic(bp, true, false); 424 425 old = xchg(&bp->xdp_prog, prog); 426 if (old) 427 bpf_prog_put(old); 428 429 if (prog) { 430 bnxt_set_rx_skb_mode(bp, true); 431 xdp_features_set_redirect_target(dev, true); 432 } else { 433 int rx, tx; 434 435 xdp_features_clear_redirect_target(dev); 436 bnxt_set_rx_skb_mode(bp, false); 437 bnxt_get_max_rings(bp, &rx, &tx, true); 438 if (rx > 1) { 439 bp->flags &= ~BNXT_FLAG_NO_AGG_RINGS; 440 bp->dev->hw_features |= NETIF_F_LRO; 441 } 442 } 443 bp->tx_nr_rings_xdp = tx_xdp; 444 bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp; 445 bp->cp_nr_rings = max_t(int, bp->tx_nr_rings, bp->rx_nr_rings); 446 bnxt_set_tpa_flags(bp); 447 bnxt_set_ring_params(bp); 448 449 if (netif_running(dev)) 450 return bnxt_open_nic(bp, true, false); 451 452 return 0; 453 } 454 455 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp) 456 { 457 struct bnxt *bp = netdev_priv(dev); 458 int rc; 459 460 switch (xdp->command) { 461 case XDP_SETUP_PROG: 462 rc = bnxt_xdp_set(bp, xdp->prog); 463 break; 464 default: 465 rc = -EINVAL; 466 break; 467 } 468 return rc; 469 } 470 471 struct sk_buff * 472 bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags, 473 struct page_pool *pool, struct xdp_buff *xdp, 474 struct rx_cmp_ext *rxcmp1) 475 { 476 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 477 478 if (!skb) 479 return NULL; 480 skb_checksum_none_assert(skb); 481 if (RX_CMP_L4_CS_OK(rxcmp1)) { 482 if (bp->dev->features & NETIF_F_RXCSUM) { 483 skb->ip_summed = CHECKSUM_UNNECESSARY; 484 skb->csum_level = RX_CMP_ENCAP(rxcmp1); 485 } 486 } 487 xdp_update_skb_shared_info(skb, num_frags, 488 sinfo->xdp_frags_size, 489 BNXT_RX_PAGE_SIZE * sinfo->nr_frags, 490 xdp_buff_is_frag_pfmemalloc(xdp)); 491 return skb; 492 } 493