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 nr_pkts) 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 struct bnxt_sw_tx_bd *tx_buf; 134 u16 tx_cons = txr->tx_cons; 135 u16 last_tx_cons = tx_cons; 136 int i, j, frags; 137 138 for (i = 0; i < nr_pkts; i++) { 139 tx_buf = &txr->tx_buf_ring[tx_cons]; 140 141 if (tx_buf->action == XDP_REDIRECT) { 142 struct pci_dev *pdev = bp->pdev; 143 144 dma_unmap_single(&pdev->dev, 145 dma_unmap_addr(tx_buf, mapping), 146 dma_unmap_len(tx_buf, len), 147 DMA_TO_DEVICE); 148 xdp_return_frame(tx_buf->xdpf); 149 tx_buf->action = 0; 150 tx_buf->xdpf = NULL; 151 } else if (tx_buf->action == XDP_TX) { 152 tx_buf->action = 0; 153 rx_doorbell_needed = true; 154 last_tx_cons = tx_cons; 155 156 frags = tx_buf->nr_frags; 157 for (j = 0; j < frags; j++) { 158 tx_cons = NEXT_TX(tx_cons); 159 tx_buf = &txr->tx_buf_ring[tx_cons]; 160 page_pool_recycle_direct(rxr->page_pool, tx_buf->page); 161 } 162 } else { 163 bnxt_sched_reset_txr(bp, txr, i); 164 return; 165 } 166 tx_cons = NEXT_TX(tx_cons); 167 } 168 WRITE_ONCE(txr->tx_cons, tx_cons); 169 if (rx_doorbell_needed) { 170 tx_buf = &txr->tx_buf_ring[last_tx_cons]; 171 bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod); 172 173 } 174 } 175 176 bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr) 177 { 178 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 179 180 return !!xdp_prog; 181 } 182 183 void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, 184 u16 cons, u8 *data_ptr, unsigned int len, 185 struct xdp_buff *xdp) 186 { 187 struct bnxt_sw_rx_bd *rx_buf; 188 u32 buflen = PAGE_SIZE; 189 struct pci_dev *pdev; 190 dma_addr_t mapping; 191 u32 offset; 192 193 pdev = bp->pdev; 194 rx_buf = &rxr->rx_buf_ring[cons]; 195 offset = bp->rx_offset; 196 197 mapping = rx_buf->mapping - bp->rx_dma_offset; 198 dma_sync_single_for_cpu(&pdev->dev, mapping + offset, len, bp->rx_dir); 199 200 xdp_init_buff(xdp, buflen, &rxr->xdp_rxq); 201 xdp_prepare_buff(xdp, data_ptr - offset, offset, len, false); 202 } 203 204 void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr, 205 struct xdp_buff *xdp) 206 { 207 struct skb_shared_info *shinfo; 208 int i; 209 210 if (!xdp || !xdp_buff_has_frags(xdp)) 211 return; 212 shinfo = xdp_get_shared_info_from_buff(xdp); 213 for (i = 0; i < shinfo->nr_frags; i++) { 214 struct page *page = skb_frag_page(&shinfo->frags[i]); 215 216 page_pool_recycle_direct(rxr->page_pool, page); 217 } 218 shinfo->nr_frags = 0; 219 } 220 221 /* returns the following: 222 * true - packet consumed by XDP and new buffer is allocated. 223 * false - packet should be passed to the stack. 224 */ 225 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons, 226 struct xdp_buff xdp, struct page *page, u8 **data_ptr, 227 unsigned int *len, u8 *event) 228 { 229 struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog); 230 struct bnxt_tx_ring_info *txr; 231 struct bnxt_sw_rx_bd *rx_buf; 232 struct pci_dev *pdev; 233 dma_addr_t mapping; 234 u32 tx_needed = 1; 235 void *orig_data; 236 u32 tx_avail; 237 u32 offset; 238 u32 act; 239 240 if (!xdp_prog) 241 return false; 242 243 pdev = bp->pdev; 244 offset = bp->rx_offset; 245 246 txr = rxr->bnapi->tx_ring; 247 /* BNXT_RX_PAGE_MODE(bp) when XDP enabled */ 248 orig_data = xdp.data; 249 250 act = bpf_prog_run_xdp(xdp_prog, &xdp); 251 252 tx_avail = bnxt_tx_avail(bp, txr); 253 /* If the tx ring is not full, we must not update the rx producer yet 254 * because we may still be transmitting on some BDs. 255 */ 256 if (tx_avail != bp->tx_ring_size) 257 *event &= ~BNXT_RX_EVENT; 258 259 *len = xdp.data_end - xdp.data; 260 if (orig_data != xdp.data) { 261 offset = xdp.data - xdp.data_hard_start; 262 *data_ptr = xdp.data_hard_start + offset; 263 } 264 265 switch (act) { 266 case XDP_PASS: 267 return false; 268 269 case XDP_TX: 270 rx_buf = &rxr->rx_buf_ring[cons]; 271 mapping = rx_buf->mapping - bp->rx_dma_offset; 272 *event = 0; 273 274 if (unlikely(xdp_buff_has_frags(&xdp))) { 275 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(&xdp); 276 277 tx_needed += sinfo->nr_frags; 278 *event = BNXT_AGG_EVENT; 279 } 280 281 if (tx_avail < tx_needed) { 282 trace_xdp_exception(bp->dev, xdp_prog, act); 283 bnxt_xdp_buff_frags_free(rxr, &xdp); 284 bnxt_reuse_rx_data(rxr, cons, page); 285 return true; 286 } 287 288 dma_sync_single_for_device(&pdev->dev, mapping + offset, *len, 289 bp->rx_dir); 290 291 *event |= BNXT_TX_EVENT; 292 __bnxt_xmit_xdp(bp, txr, mapping + offset, *len, 293 NEXT_RX(rxr->rx_prod), &xdp); 294 bnxt_reuse_rx_data(rxr, cons, page); 295 return true; 296 case XDP_REDIRECT: 297 /* if we are calling this here then we know that the 298 * redirect is coming from a frame received by the 299 * bnxt_en driver. 300 */ 301 rx_buf = &rxr->rx_buf_ring[cons]; 302 mapping = rx_buf->mapping - bp->rx_dma_offset; 303 dma_unmap_page_attrs(&pdev->dev, mapping, 304 PAGE_SIZE, bp->rx_dir, 305 DMA_ATTR_WEAK_ORDERING); 306 307 /* if we are unable to allocate a new buffer, abort and reuse */ 308 if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) { 309 trace_xdp_exception(bp->dev, xdp_prog, act); 310 bnxt_xdp_buff_frags_free(rxr, &xdp); 311 bnxt_reuse_rx_data(rxr, cons, page); 312 return true; 313 } 314 315 if (xdp_do_redirect(bp->dev, &xdp, xdp_prog)) { 316 trace_xdp_exception(bp->dev, xdp_prog, act); 317 page_pool_recycle_direct(rxr->page_pool, page); 318 return true; 319 } 320 321 *event |= BNXT_REDIRECT_EVENT; 322 break; 323 default: 324 bpf_warn_invalid_xdp_action(bp->dev, xdp_prog, act); 325 fallthrough; 326 case XDP_ABORTED: 327 trace_xdp_exception(bp->dev, xdp_prog, act); 328 fallthrough; 329 case XDP_DROP: 330 bnxt_xdp_buff_frags_free(rxr, &xdp); 331 bnxt_reuse_rx_data(rxr, cons, page); 332 break; 333 } 334 return true; 335 } 336 337 int bnxt_xdp_xmit(struct net_device *dev, int num_frames, 338 struct xdp_frame **frames, u32 flags) 339 { 340 struct bnxt *bp = netdev_priv(dev); 341 struct bpf_prog *xdp_prog = READ_ONCE(bp->xdp_prog); 342 struct pci_dev *pdev = bp->pdev; 343 struct bnxt_tx_ring_info *txr; 344 dma_addr_t mapping; 345 int nxmit = 0; 346 int ring; 347 int i; 348 349 if (!test_bit(BNXT_STATE_OPEN, &bp->state) || 350 !bp->tx_nr_rings_xdp || 351 !xdp_prog) 352 return -EINVAL; 353 354 ring = smp_processor_id() % bp->tx_nr_rings_xdp; 355 txr = &bp->tx_ring[ring]; 356 357 if (READ_ONCE(txr->dev_state) == BNXT_DEV_STATE_CLOSING) 358 return -EINVAL; 359 360 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 361 spin_lock(&txr->xdp_tx_lock); 362 363 for (i = 0; i < num_frames; i++) { 364 struct xdp_frame *xdp = frames[i]; 365 366 if (!bnxt_tx_avail(bp, txr)) 367 break; 368 369 mapping = dma_map_single(&pdev->dev, xdp->data, xdp->len, 370 DMA_TO_DEVICE); 371 372 if (dma_mapping_error(&pdev->dev, mapping)) 373 break; 374 375 __bnxt_xmit_xdp_redirect(bp, txr, mapping, xdp->len, xdp); 376 nxmit++; 377 } 378 379 if (flags & XDP_XMIT_FLUSH) { 380 /* Sync BD data before updating doorbell */ 381 wmb(); 382 bnxt_db_write(bp, &txr->tx_db, txr->tx_prod); 383 } 384 385 if (static_branch_unlikely(&bnxt_xdp_locking_key)) 386 spin_unlock(&txr->xdp_tx_lock); 387 388 return nxmit; 389 } 390 391 /* Under rtnl_lock */ 392 static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog) 393 { 394 struct net_device *dev = bp->dev; 395 int tx_xdp = 0, rc, tc; 396 struct bpf_prog *old; 397 398 if (prog && !prog->aux->xdp_has_frags && 399 bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) { 400 netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n", 401 bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU); 402 return -EOPNOTSUPP; 403 } 404 if (!(bp->flags & BNXT_FLAG_SHARED_RINGS)) { 405 netdev_warn(dev, "ethtool rx/tx channels must be combined to support XDP.\n"); 406 return -EOPNOTSUPP; 407 } 408 if (prog) 409 tx_xdp = bp->rx_nr_rings; 410 411 tc = netdev_get_num_tc(dev); 412 if (!tc) 413 tc = 1; 414 rc = bnxt_check_rings(bp, bp->tx_nr_rings_per_tc, bp->rx_nr_rings, 415 true, tc, tx_xdp); 416 if (rc) { 417 netdev_warn(dev, "Unable to reserve enough TX rings to support XDP.\n"); 418 return rc; 419 } 420 if (netif_running(dev)) 421 bnxt_close_nic(bp, true, false); 422 423 old = xchg(&bp->xdp_prog, prog); 424 if (old) 425 bpf_prog_put(old); 426 427 if (prog) { 428 bnxt_set_rx_skb_mode(bp, true); 429 xdp_features_set_redirect_target(dev, true); 430 } else { 431 int rx, tx; 432 433 xdp_features_clear_redirect_target(dev); 434 bnxt_set_rx_skb_mode(bp, false); 435 bnxt_get_max_rings(bp, &rx, &tx, true); 436 if (rx > 1) { 437 bp->flags &= ~BNXT_FLAG_NO_AGG_RINGS; 438 bp->dev->hw_features |= NETIF_F_LRO; 439 } 440 } 441 bp->tx_nr_rings_xdp = tx_xdp; 442 bp->tx_nr_rings = bp->tx_nr_rings_per_tc * tc + tx_xdp; 443 bp->cp_nr_rings = max_t(int, bp->tx_nr_rings, bp->rx_nr_rings); 444 bnxt_set_tpa_flags(bp); 445 bnxt_set_ring_params(bp); 446 447 if (netif_running(dev)) 448 return bnxt_open_nic(bp, true, false); 449 450 return 0; 451 } 452 453 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp) 454 { 455 struct bnxt *bp = netdev_priv(dev); 456 int rc; 457 458 switch (xdp->command) { 459 case XDP_SETUP_PROG: 460 rc = bnxt_xdp_set(bp, xdp->prog); 461 break; 462 default: 463 rc = -EINVAL; 464 break; 465 } 466 return rc; 467 } 468 469 struct sk_buff * 470 bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags, 471 struct page_pool *pool, struct xdp_buff *xdp, 472 struct rx_cmp_ext *rxcmp1) 473 { 474 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 475 476 if (!skb) 477 return NULL; 478 skb_checksum_none_assert(skb); 479 if (RX_CMP_L4_CS_OK(rxcmp1)) { 480 if (bp->dev->features & NETIF_F_RXCSUM) { 481 skb->ip_summed = CHECKSUM_UNNECESSARY; 482 skb->csum_level = RX_CMP_ENCAP(rxcmp1); 483 } 484 } 485 xdp_update_skb_shared_info(skb, num_frags, 486 sinfo->xdp_frags_size, 487 PAGE_SIZE * sinfo->nr_frags, 488 xdp_buff_is_frag_pfmemalloc(xdp)); 489 return skb; 490 } 491