ethernet/sfc/rx.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /****************************************************************************
0003  * Driver for Solarflare network controllers and boards
0004  * Copyright 2005-2006 Fen Systems Ltd.
0005  * Copyright 2005-2013 Solarflare Communications Inc.
0006  */
0007
0008 #include <linux/socket.h>
0009 #include <linux/in.h>
0010 #include <linux/slab.h>
0011 #include <linux/ip.h>
0012 #include <linux/ipv6.h>
0013 #include <linux/tcp.h>
0014 #include <linux/udp.h>
0015 #include <linux/prefetch.h>
0016 #include <linux/moduleparam.h>
0017 #include <linux/iommu.h>
0018 #include <net/ip.h>
0019 #include <net/checksum.h>
0020 #include <net/xdp.h>
0021 #include <linux/bpf_trace.h>
0022 #include "net_driver.h"
0023 #include "efx.h"
0024 #include "rx_common.h"
0025 #include "filter.h"
0026 #include "nic.h"
0027 #include "selftest.h"
0028 #include "workarounds.h"
0029
0030 /* Preferred number of descriptors to fill at once */
0031 #define EFX_RX_PREFERRED_BATCH 8U
0032
0033 /* Maximum rx prefix used by any architecture. */
0034 #define EFX_MAX_RX_PREFIX_SIZE 16
0035
0036 /* Size of buffer allocated for skb header area. */
0037 #define EFX_SKB_HEADERS  128u
0038
0039 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
0040 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
0041                       EFX_RX_USR_BUF_SIZE)
0042
0043 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
0044                      struct efx_rx_buffer *rx_buf,
0045                      int len)
0046 {
0047     struct efx_nic *efx = rx_queue->efx;
0048     unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
0049
0050     if (likely(len <= max_len))
0051         return;
0052
0053     /* The packet must be discarded, but this is only a fatal error
0054      * if the caller indicated it was
0055      */
0056     rx_buf->flags |= EFX_RX_PKT_DISCARD;
0057
0058     if (net_ratelimit())
0059         netif_err(efx, rx_err, efx->net_dev,
0060               "RX queue %d overlength RX event (%#x > %#x)\n",
0061               efx_rx_queue_index(rx_queue), len, max_len);
0062
0063     efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
0064 }
0065
0066 /* Allocate and construct an SKB around page fragments */
0067 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
0068                      struct efx_rx_buffer *rx_buf,
0069                      unsigned int n_frags,
0070                      u8 *eh, int hdr_len)
0071 {
0072     struct efx_nic *efx = channel->efx;
0073     struct sk_buff *skb;
0074
0075     /* Allocate an SKB to store the headers */
0076     skb = netdev_alloc_skb(efx->net_dev,
0077                    efx->rx_ip_align + efx->rx_prefix_size +
0078                    hdr_len);
0079     if (unlikely(skb == NULL)) {
0080         atomic_inc(&efx->n_rx_noskb_drops);
0081         return NULL;
0082     }
0083
0084     EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
0085
0086     memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
0087            efx->rx_prefix_size + hdr_len);
0088     skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
0089     __skb_put(skb, hdr_len);
0090
0091     /* Append the remaining page(s) onto the frag list */
0092     if (rx_buf->len > hdr_len) {
0093         rx_buf->page_offset += hdr_len;
0094         rx_buf->len -= hdr_len;
0095
0096         for (;;) {
0097             skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
0098                     rx_buf->page, rx_buf->page_offset,
0099                     rx_buf->len, efx->rx_buffer_truesize);
0100             rx_buf->page = NULL;
0101
0102             if (skb_shinfo(skb)->nr_frags == n_frags)
0103                 break;
0104
0105             rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
0106         }
0107     } else {
0108         __free_pages(rx_buf->page, efx->rx_buffer_order);
0109         rx_buf->page = NULL;
0110         n_frags = 0;
0111     }
0112
0113     /* Move past the ethernet header */
0114     skb->protocol = eth_type_trans(skb, efx->net_dev);
0115
0116     skb_mark_napi_id(skb, &channel->napi_str);
0117
0118     return skb;
0119 }
0120
0121 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
0122            unsigned int n_frags, unsigned int len, u16 flags)
0123 {
0124     struct efx_nic *efx = rx_queue->efx;
0125     struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
0126     struct efx_rx_buffer *rx_buf;
0127
0128     rx_queue->rx_packets++;
0129
0130     rx_buf = efx_rx_buffer(rx_queue, index);
0131     rx_buf->flags |= flags;
0132
0133     /* Validate the number of fragments and completed length */
0134     if (n_frags == 1) {
0135         if (!(flags & EFX_RX_PKT_PREFIX_LEN))
0136             efx_rx_packet__check_len(rx_queue, rx_buf, len);
0137     } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
0138            unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
0139            unlikely(len > n_frags * efx->rx_dma_len) ||
0140            unlikely(!efx->rx_scatter)) {
0141         /* If this isn't an explicit discard request, either
0142          * the hardware or the driver is broken.
0143          */
0144         WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
0145         rx_buf->flags |= EFX_RX_PKT_DISCARD;
0146     }
0147
0148     netif_vdbg(efx, rx_status, efx->net_dev,
0149            "RX queue %d received ids %x-%x len %d %s%s\n",
0150            efx_rx_queue_index(rx_queue), index,
0151            (index + n_frags - 1) & rx_queue->ptr_mask, len,
0152            (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
0153            (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
0154
0155     /* Discard packet, if instructed to do so.  Process the
0156      * previous receive first.
0157      */
0158     if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
0159         efx_rx_flush_packet(channel);
0160         efx_discard_rx_packet(channel, rx_buf, n_frags);
0161         return;
0162     }
0163
0164     if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
0165         rx_buf->len = len;
0166
0167     /* Release and/or sync the DMA mapping - assumes all RX buffers
0168      * consumed in-order per RX queue.
0169      */
0170     efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
0171
0172     /* Prefetch nice and early so data will (hopefully) be in cache by
0173      * the time we look at it.
0174      */
0175     prefetch(efx_rx_buf_va(rx_buf));
0176
0177     rx_buf->page_offset += efx->rx_prefix_size;
0178     rx_buf->len -= efx->rx_prefix_size;
0179
0180     if (n_frags > 1) {
0181         /* Release/sync DMA mapping for additional fragments.
0182          * Fix length for last fragment.
0183          */
0184         unsigned int tail_frags = n_frags - 1;
0185
0186         for (;;) {
0187             rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
0188             if (--tail_frags == 0)
0189                 break;
0190             efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
0191         }
0192         rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
0193         efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
0194     }
0195
0196     /* All fragments have been DMA-synced, so recycle pages. */
0197     rx_buf = efx_rx_buffer(rx_queue, index);
0198     efx_recycle_rx_pages(channel, rx_buf, n_frags);
0199
0200     /* Pipeline receives so that we give time for packet headers to be
0201      * prefetched into cache.
0202      */
0203     efx_rx_flush_packet(channel);
0204     channel->rx_pkt_n_frags = n_frags;
0205     channel->rx_pkt_index = index;
0206 }
0207
0208 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
0209                struct efx_rx_buffer *rx_buf,
0210                unsigned int n_frags)
0211 {
0212     struct sk_buff *skb;
0213     u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
0214
0215     skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
0216     if (unlikely(skb == NULL)) {
0217         struct efx_rx_queue *rx_queue;
0218
0219         rx_queue = efx_channel_get_rx_queue(channel);
0220         efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
0221         return;
0222     }
0223     skb_record_rx_queue(skb, channel->rx_queue.core_index);
0224
0225     /* Set the SKB flags */
0226     skb_checksum_none_assert(skb);
0227     if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
0228         skb->ip_summed = CHECKSUM_UNNECESSARY;
0229         skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
0230     }
0231
0232     efx_rx_skb_attach_timestamp(channel, skb);
0233
0234     if (channel->type->receive_skb)
0235         if (channel->type->receive_skb(channel, skb))
0236             return;
0237
0238     /* Pass the packet up */
0239     if (channel->rx_list != NULL)
0240         /* Add to list, will pass up later */
0241         list_add_tail(&skb->list, channel->rx_list);
0242     else
0243         /* No list, so pass it up now */
0244         netif_receive_skb(skb);
0245 }
0246
0247 /** efx_do_xdp: perform XDP processing on a received packet
0248  *
0249  * Returns true if packet should still be delivered.
0250  */
0251 static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
0252                struct efx_rx_buffer *rx_buf, u8 **ehp)
0253 {
0254     u8 rx_prefix[EFX_MAX_RX_PREFIX_SIZE];
0255     struct efx_rx_queue *rx_queue;
0256     struct bpf_prog *xdp_prog;
0257     struct xdp_frame *xdpf;
0258     struct xdp_buff xdp;
0259     u32 xdp_act;
0260     s16 offset;
0261     int err;
0262
0263     xdp_prog = rcu_dereference_bh(efx->xdp_prog);
0264     if (!xdp_prog)
0265         return true;
0266
0267     rx_queue = efx_channel_get_rx_queue(channel);
0268
0269     if (unlikely(channel->rx_pkt_n_frags > 1)) {
0270         /* We can't do XDP on fragmented packets - drop. */
0271         efx_free_rx_buffers(rx_queue, rx_buf,
0272                     channel->rx_pkt_n_frags);
0273         if (net_ratelimit())
0274             netif_err(efx, rx_err, efx->net_dev,
0275                   "XDP is not possible with multiple receive fragments (%d)\n",
0276                   channel->rx_pkt_n_frags);
0277         channel->n_rx_xdp_bad_drops++;
0278         return false;
0279     }
0280
0281     dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr,
0282                 rx_buf->len, DMA_FROM_DEVICE);
0283
0284     /* Save the rx prefix. */
0285     EFX_WARN_ON_PARANOID(efx->rx_prefix_size > EFX_MAX_RX_PREFIX_SIZE);
0286     memcpy(rx_prefix, *ehp - efx->rx_prefix_size,
0287            efx->rx_prefix_size);
0288
0289     xdp_init_buff(&xdp, efx->rx_page_buf_step, &rx_queue->xdp_rxq_info);
0290     /* No support yet for XDP metadata */
0291     xdp_prepare_buff(&xdp, *ehp - EFX_XDP_HEADROOM, EFX_XDP_HEADROOM,
0292              rx_buf->len, false);
0293
0294     xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
0295
0296     offset = (u8 *)xdp.data - *ehp;
0297
0298     switch (xdp_act) {
0299     case XDP_PASS:
0300         /* Fix up rx prefix. */
0301         if (offset) {
0302             *ehp += offset;
0303             rx_buf->page_offset += offset;
0304             rx_buf->len -= offset;
0305             memcpy(*ehp - efx->rx_prefix_size, rx_prefix,
0306                    efx->rx_prefix_size);
0307         }
0308         break;
0309
0310     case XDP_TX:
0311         /* Buffer ownership passes to tx on success. */
0312         xdpf = xdp_convert_buff_to_frame(&xdp);
0313         err = efx_xdp_tx_buffers(efx, 1, &xdpf, true);
0314         if (unlikely(err != 1)) {
0315             efx_free_rx_buffers(rx_queue, rx_buf, 1);
0316             if (net_ratelimit())
0317                 netif_err(efx, rx_err, efx->net_dev,
0318                       "XDP TX failed (%d)\n", err);
0319             channel->n_rx_xdp_bad_drops++;
0320             trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
0321         } else {
0322             channel->n_rx_xdp_tx++;
0323         }
0324         break;
0325
0326     case XDP_REDIRECT:
0327         err = xdp_do_redirect(efx->net_dev, &xdp, xdp_prog);
0328         if (unlikely(err)) {
0329             efx_free_rx_buffers(rx_queue, rx_buf, 1);
0330             if (net_ratelimit())
0331                 netif_err(efx, rx_err, efx->net_dev,
0332                       "XDP redirect failed (%d)\n", err);
0333             channel->n_rx_xdp_bad_drops++;
0334             trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
0335         } else {
0336             channel->n_rx_xdp_redirect++;
0337         }
0338         break;
0339
0340     default:
0341         bpf_warn_invalid_xdp_action(efx->net_dev, xdp_prog, xdp_act);
0342         efx_free_rx_buffers(rx_queue, rx_buf, 1);
0343         channel->n_rx_xdp_bad_drops++;
0344         trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
0345         break;
0346
0347     case XDP_ABORTED:
0348         trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
0349         fallthrough;
0350     case XDP_DROP:
0351         efx_free_rx_buffers(rx_queue, rx_buf, 1);
0352         channel->n_rx_xdp_drops++;
0353         break;
0354     }
0355
0356     return xdp_act == XDP_PASS;
0357 }
0358
0359 /* Handle a received packet.  Second half: Touches packet payload. */
0360 void __efx_rx_packet(struct efx_channel *channel)
0361 {
0362     struct efx_nic *efx = channel->efx;
0363     struct efx_rx_buffer *rx_buf =
0364         efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
0365     u8 *eh = efx_rx_buf_va(rx_buf);
0366
0367     /* Read length from the prefix if necessary.  This already
0368      * excludes the length of the prefix itself.
0369      */
0370     if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
0371         rx_buf->len = le16_to_cpup((__le16 *)
0372                        (eh + efx->rx_packet_len_offset));
0373
0374     /* If we're in loopback test, then pass the packet directly to the
0375      * loopback layer, and free the rx_buf here
0376      */
0377     if (unlikely(efx->loopback_selftest)) {
0378         struct efx_rx_queue *rx_queue;
0379
0380         efx_loopback_rx_packet(efx, eh, rx_buf->len);
0381         rx_queue = efx_channel_get_rx_queue(channel);
0382         efx_free_rx_buffers(rx_queue, rx_buf,
0383                     channel->rx_pkt_n_frags);
0384         goto out;
0385     }
0386
0387     if (!efx_do_xdp(efx, channel, rx_buf, &eh))
0388         goto out;
0389
0390     if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
0391         rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
0392
0393     if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
0394         efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh, 0);
0395     else
0396         efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
0397 out:
0398     channel->rx_pkt_n_frags = 0;
0399 }