1 // SPDX-License-Identifier: GPL-2.0-only 2 /**************************************************************************** 3 * Driver for Solarflare network controllers and boards 4 * Copyright 2005-2006 Fen Systems Ltd. 5 * Copyright 2005-2013 Solarflare Communications Inc. 6 */ 7 8 #include <linux/socket.h> 9 #include <linux/in.h> 10 #include <linux/slab.h> 11 #include <linux/ip.h> 12 #include <linux/ipv6.h> 13 #include <linux/tcp.h> 14 #include <linux/udp.h> 15 #include <linux/prefetch.h> 16 #include <linux/moduleparam.h> 17 #include <linux/iommu.h> 18 #include <net/ip.h> 19 #include <net/checksum.h> 20 #include <net/xdp.h> 21 #include <linux/bpf_trace.h> 22 #include "net_driver.h" 23 #include "efx.h" 24 #include "rx_common.h" 25 #include "filter.h" 26 #include "nic.h" 27 #include "selftest.h" 28 #include "workarounds.h" 29 30 /* Preferred number of descriptors to fill at once */ 31 #define EFX_RX_PREFERRED_BATCH 8U 32 33 /* Maximum rx prefix used by any architecture. */ 34 #define EFX_MAX_RX_PREFIX_SIZE 16 35 36 /* Size of buffer allocated for skb header area. */ 37 #define EFX_SKB_HEADERS 128u 38 39 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */ 40 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \ 41 EFX_RX_USR_BUF_SIZE) 42 43 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue, 44 struct efx_rx_buffer *rx_buf, 45 int len) 46 { 47 struct efx_nic *efx = rx_queue->efx; 48 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding; 49 50 if (likely(len <= max_len)) 51 return; 52 53 /* The packet must be discarded, but this is only a fatal error 54 * if the caller indicated it was 55 */ 56 rx_buf->flags |= EFX_RX_PKT_DISCARD; 57 58 if (net_ratelimit()) 59 netif_err(efx, rx_err, efx->net_dev, 60 "RX queue %d overlength RX event (%#x > %#x)\n", 61 efx_rx_queue_index(rx_queue), len, max_len); 62 63 efx_rx_queue_channel(rx_queue)->n_rx_overlength++; 64 } 65 66 /* Allocate and construct an SKB around page fragments */ 67 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel, 68 struct efx_rx_buffer *rx_buf, 69 unsigned int n_frags, 70 u8 *eh, int hdr_len) 71 { 72 struct efx_nic *efx = channel->efx; 73 struct sk_buff *skb; 74 75 /* Allocate an SKB to store the headers */ 76 skb = netdev_alloc_skb(efx->net_dev, 77 efx->rx_ip_align + efx->rx_prefix_size + 78 hdr_len); 79 if (unlikely(skb == NULL)) { 80 atomic_inc(&efx->n_rx_noskb_drops); 81 return NULL; 82 } 83 84 EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len); 85 86 memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size, 87 efx->rx_prefix_size + hdr_len); 88 skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size); 89 __skb_put(skb, hdr_len); 90 91 /* Append the remaining page(s) onto the frag list */ 92 if (rx_buf->len > hdr_len) { 93 rx_buf->page_offset += hdr_len; 94 rx_buf->len -= hdr_len; 95 96 for (;;) { 97 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, 98 rx_buf->page, rx_buf->page_offset, 99 rx_buf->len, efx->rx_buffer_truesize); 100 rx_buf->page = NULL; 101 102 if (skb_shinfo(skb)->nr_frags == n_frags) 103 break; 104 105 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf); 106 } 107 } else { 108 __free_pages(rx_buf->page, efx->rx_buffer_order); 109 rx_buf->page = NULL; 110 n_frags = 0; 111 } 112 113 /* Move past the ethernet header */ 114 skb->protocol = eth_type_trans(skb, efx->net_dev); 115 116 skb_mark_napi_id(skb, &channel->napi_str); 117 118 return skb; 119 } 120 121 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index, 122 unsigned int n_frags, unsigned int len, u16 flags) 123 { 124 struct efx_nic *efx = rx_queue->efx; 125 struct efx_channel *channel = efx_rx_queue_channel(rx_queue); 126 struct efx_rx_buffer *rx_buf; 127 128 rx_buf = efx_rx_buffer(rx_queue, index); 129 rx_buf->flags |= flags; 130 131 /* Validate the number of fragments and completed length */ 132 if (n_frags == 1) { 133 if (!(flags & EFX_RX_PKT_PREFIX_LEN)) 134 efx_rx_packet__check_len(rx_queue, rx_buf, len); 135 } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) || 136 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) || 137 unlikely(len > n_frags * efx->rx_dma_len) || 138 unlikely(!efx->rx_scatter)) { 139 /* If this isn't an explicit discard request, either 140 * the hardware or the driver is broken. 141 */ 142 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD)); 143 rx_buf->flags |= EFX_RX_PKT_DISCARD; 144 } 145 146 netif_vdbg(efx, rx_status, efx->net_dev, 147 "RX queue %d received ids %x-%x len %d %s%s\n", 148 efx_rx_queue_index(rx_queue), index, 149 (index + n_frags - 1) & rx_queue->ptr_mask, len, 150 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "", 151 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : ""); 152 153 /* Discard packet, if instructed to do so. Process the 154 * previous receive first. 155 */ 156 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) { 157 efx_rx_flush_packet(channel); 158 efx_discard_rx_packet(channel, rx_buf, n_frags); 159 return; 160 } 161 162 if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN)) 163 rx_buf->len = len; 164 165 /* Release and/or sync the DMA mapping - assumes all RX buffers 166 * consumed in-order per RX queue. 167 */ 168 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len); 169 170 /* Prefetch nice and early so data will (hopefully) be in cache by 171 * the time we look at it. 172 */ 173 prefetch(efx_rx_buf_va(rx_buf)); 174 175 rx_buf->page_offset += efx->rx_prefix_size; 176 rx_buf->len -= efx->rx_prefix_size; 177 178 if (n_frags > 1) { 179 /* Release/sync DMA mapping for additional fragments. 180 * Fix length for last fragment. 181 */ 182 unsigned int tail_frags = n_frags - 1; 183 184 for (;;) { 185 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 186 if (--tail_frags == 0) 187 break; 188 efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len); 189 } 190 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len; 191 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len); 192 } 193 194 /* All fragments have been DMA-synced, so recycle pages. */ 195 rx_buf = efx_rx_buffer(rx_queue, index); 196 efx_recycle_rx_pages(channel, rx_buf, n_frags); 197 198 /* Pipeline receives so that we give time for packet headers to be 199 * prefetched into cache. 200 */ 201 efx_rx_flush_packet(channel); 202 channel->rx_pkt_n_frags = n_frags; 203 channel->rx_pkt_index = index; 204 } 205 206 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh, 207 struct efx_rx_buffer *rx_buf, 208 unsigned int n_frags) 209 { 210 struct sk_buff *skb; 211 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS); 212 213 skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len); 214 if (unlikely(skb == NULL)) { 215 struct efx_rx_queue *rx_queue; 216 217 rx_queue = efx_channel_get_rx_queue(channel); 218 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 219 return; 220 } 221 skb_record_rx_queue(skb, channel->rx_queue.core_index); 222 223 /* Set the SKB flags */ 224 skb_checksum_none_assert(skb); 225 if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) { 226 skb->ip_summed = CHECKSUM_UNNECESSARY; 227 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL); 228 } 229 230 efx_rx_skb_attach_timestamp(channel, skb); 231 232 if (channel->type->receive_skb) 233 if (channel->type->receive_skb(channel, skb)) 234 return; 235 236 /* Pass the packet up */ 237 if (channel->rx_list != NULL) 238 /* Add to list, will pass up later */ 239 list_add_tail(&skb->list, channel->rx_list); 240 else 241 /* No list, so pass it up now */ 242 netif_receive_skb(skb); 243 } 244 245 /** efx_do_xdp: perform XDP processing on a received packet 246 * 247 * Returns true if packet should still be delivered. 248 */ 249 static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel, 250 struct efx_rx_buffer *rx_buf, u8 **ehp) 251 { 252 u8 rx_prefix[EFX_MAX_RX_PREFIX_SIZE]; 253 struct efx_rx_queue *rx_queue; 254 struct bpf_prog *xdp_prog; 255 struct xdp_frame *xdpf; 256 struct xdp_buff xdp; 257 u32 xdp_act; 258 s16 offset; 259 int err; 260 261 xdp_prog = rcu_dereference_bh(efx->xdp_prog); 262 if (!xdp_prog) 263 return true; 264 265 rx_queue = efx_channel_get_rx_queue(channel); 266 267 if (unlikely(channel->rx_pkt_n_frags > 1)) { 268 /* We can't do XDP on fragmented packets - drop. */ 269 efx_free_rx_buffers(rx_queue, rx_buf, 270 channel->rx_pkt_n_frags); 271 if (net_ratelimit()) 272 netif_err(efx, rx_err, efx->net_dev, 273 "XDP is not possible with multiple receive fragments (%d)\n", 274 channel->rx_pkt_n_frags); 275 channel->n_rx_xdp_bad_drops++; 276 return false; 277 } 278 279 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, 280 rx_buf->len, DMA_FROM_DEVICE); 281 282 /* Save the rx prefix. */ 283 EFX_WARN_ON_PARANOID(efx->rx_prefix_size > EFX_MAX_RX_PREFIX_SIZE); 284 memcpy(rx_prefix, *ehp - efx->rx_prefix_size, 285 efx->rx_prefix_size); 286 287 xdp_init_buff(&xdp, efx->rx_page_buf_step, &rx_queue->xdp_rxq_info); 288 /* No support yet for XDP metadata */ 289 xdp_prepare_buff(&xdp, *ehp - EFX_XDP_HEADROOM, EFX_XDP_HEADROOM, 290 rx_buf->len, false); 291 292 xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp); 293 294 offset = (u8 *)xdp.data - *ehp; 295 296 switch (xdp_act) { 297 case XDP_PASS: 298 /* Fix up rx prefix. */ 299 if (offset) { 300 *ehp += offset; 301 rx_buf->page_offset += offset; 302 rx_buf->len -= offset; 303 memcpy(*ehp - efx->rx_prefix_size, rx_prefix, 304 efx->rx_prefix_size); 305 } 306 break; 307 308 case XDP_TX: 309 /* Buffer ownership passes to tx on success. */ 310 xdpf = xdp_convert_buff_to_frame(&xdp); 311 err = efx_xdp_tx_buffers(efx, 1, &xdpf, true); 312 if (unlikely(err != 1)) { 313 efx_free_rx_buffers(rx_queue, rx_buf, 1); 314 if (net_ratelimit()) 315 netif_err(efx, rx_err, efx->net_dev, 316 "XDP TX failed (%d)\n", err); 317 channel->n_rx_xdp_bad_drops++; 318 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act); 319 } else { 320 channel->n_rx_xdp_tx++; 321 } 322 break; 323 324 case XDP_REDIRECT: 325 err = xdp_do_redirect(efx->net_dev, &xdp, xdp_prog); 326 if (unlikely(err)) { 327 efx_free_rx_buffers(rx_queue, rx_buf, 1); 328 if (net_ratelimit()) 329 netif_err(efx, rx_err, efx->net_dev, 330 "XDP redirect failed (%d)\n", err); 331 channel->n_rx_xdp_bad_drops++; 332 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act); 333 } else { 334 channel->n_rx_xdp_redirect++; 335 } 336 break; 337 338 default: 339 bpf_warn_invalid_xdp_action(efx->net_dev, xdp_prog, xdp_act); 340 efx_free_rx_buffers(rx_queue, rx_buf, 1); 341 channel->n_rx_xdp_bad_drops++; 342 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act); 343 break; 344 345 case XDP_ABORTED: 346 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act); 347 fallthrough; 348 case XDP_DROP: 349 efx_free_rx_buffers(rx_queue, rx_buf, 1); 350 channel->n_rx_xdp_drops++; 351 break; 352 } 353 354 return xdp_act == XDP_PASS; 355 } 356 357 /* Handle a received packet. Second half: Touches packet payload. */ 358 void __efx_rx_packet(struct efx_channel *channel) 359 { 360 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 361 struct efx_nic *efx = channel->efx; 362 struct efx_rx_buffer *rx_buf = 363 efx_rx_buffer(rx_queue, channel->rx_pkt_index); 364 u8 *eh = efx_rx_buf_va(rx_buf); 365 366 /* Read length from the prefix if necessary. This already 367 * excludes the length of the prefix itself. 368 */ 369 if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN) { 370 rx_buf->len = le16_to_cpup((__le16 *) 371 (eh + efx->rx_packet_len_offset)); 372 /* A known issue may prevent this being filled in; 373 * if that happens, just drop the packet. 374 * Must do that in the driver since passing a zero-length 375 * packet up to the stack may cause a crash. 376 */ 377 if (unlikely(!rx_buf->len)) { 378 efx_free_rx_buffers(rx_queue, rx_buf, 379 channel->rx_pkt_n_frags); 380 channel->n_rx_frm_trunc++; 381 goto out; 382 } 383 } 384 385 /* If we're in loopback test, then pass the packet directly to the 386 * loopback layer, and free the rx_buf here 387 */ 388 if (unlikely(efx->loopback_selftest)) { 389 efx_loopback_rx_packet(efx, eh, rx_buf->len); 390 efx_free_rx_buffers(rx_queue, rx_buf, 391 channel->rx_pkt_n_frags); 392 goto out; 393 } 394 395 rx_queue->rx_packets++; 396 rx_queue->rx_bytes += rx_buf->len; 397 398 if (!efx_do_xdp(efx, channel, rx_buf, &eh)) 399 goto out; 400 401 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM))) 402 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED; 403 404 if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb) 405 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh, 0); 406 else 407 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags); 408 out: 409 channel->rx_pkt_n_frags = 0; 410 } 411