Loading...
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
43static 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 */
67static 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
121void 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_queue->rx_packets++;
129
130 rx_buf = efx_rx_buffer(rx_queue, index);
131 rx_buf->flags |= flags;
132
133 /* Validate the number of fragments and completed length */
134 if (n_frags == 1) {
135 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
136 efx_rx_packet__check_len(rx_queue, rx_buf, len);
137 } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
138 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
139 unlikely(len > n_frags * efx->rx_dma_len) ||
140 unlikely(!efx->rx_scatter)) {
141 /* If this isn't an explicit discard request, either
142 * the hardware or the driver is broken.
143 */
144 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
145 rx_buf->flags |= EFX_RX_PKT_DISCARD;
146 }
147
148 netif_vdbg(efx, rx_status, efx->net_dev,
149 "RX queue %d received ids %x-%x len %d %s%s\n",
150 efx_rx_queue_index(rx_queue), index,
151 (index + n_frags - 1) & rx_queue->ptr_mask, len,
152 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
153 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
154
155 /* Discard packet, if instructed to do so. Process the
156 * previous receive first.
157 */
158 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
159 efx_rx_flush_packet(channel);
160 efx_discard_rx_packet(channel, rx_buf, n_frags);
161 return;
162 }
163
164 if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
165 rx_buf->len = len;
166
167 /* Release and/or sync the DMA mapping - assumes all RX buffers
168 * consumed in-order per RX queue.
169 */
170 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
171
172 /* Prefetch nice and early so data will (hopefully) be in cache by
173 * the time we look at it.
174 */
175 prefetch(efx_rx_buf_va(rx_buf));
176
177 rx_buf->page_offset += efx->rx_prefix_size;
178 rx_buf->len -= efx->rx_prefix_size;
179
180 if (n_frags > 1) {
181 /* Release/sync DMA mapping for additional fragments.
182 * Fix length for last fragment.
183 */
184 unsigned int tail_frags = n_frags - 1;
185
186 for (;;) {
187 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
188 if (--tail_frags == 0)
189 break;
190 efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
191 }
192 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
193 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
194 }
195
196 /* All fragments have been DMA-synced, so recycle pages. */
197 rx_buf = efx_rx_buffer(rx_queue, index);
198 efx_recycle_rx_pages(channel, rx_buf, n_frags);
199
200 /* Pipeline receives so that we give time for packet headers to be
201 * prefetched into cache.
202 */
203 efx_rx_flush_packet(channel);
204 channel->rx_pkt_n_frags = n_frags;
205 channel->rx_pkt_index = index;
206}
207
208static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
209 struct efx_rx_buffer *rx_buf,
210 unsigned int n_frags)
211{
212 struct sk_buff *skb;
213 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
214
215 skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
216 if (unlikely(skb == NULL)) {
217 struct efx_rx_queue *rx_queue;
218
219 rx_queue = efx_channel_get_rx_queue(channel);
220 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
221 return;
222 }
223 skb_record_rx_queue(skb, channel->rx_queue.core_index);
224
225 /* Set the SKB flags */
226 skb_checksum_none_assert(skb);
227 if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
228 skb->ip_summed = CHECKSUM_UNNECESSARY;
229 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
230 }
231
232 efx_rx_skb_attach_timestamp(channel, skb);
233
234 if (channel->type->receive_skb)
235 if (channel->type->receive_skb(channel, skb))
236 return;
237
238 /* Pass the packet up */
239 if (channel->rx_list != NULL)
240 /* Add to list, will pass up later */
241 list_add_tail(&skb->list, channel->rx_list);
242 else
243 /* No list, so pass it up now */
244 netif_receive_skb(skb);
245}
246
247/** efx_do_xdp: perform XDP processing on a received packet
248 *
249 * Returns true if packet should still be delivered.
250 */
251static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
252 struct efx_rx_buffer *rx_buf, u8 **ehp)
253{
254 u8 rx_prefix[EFX_MAX_RX_PREFIX_SIZE];
255 struct efx_rx_queue *rx_queue;
256 struct bpf_prog *xdp_prog;
257 struct xdp_frame *xdpf;
258 struct xdp_buff xdp;
259 u32 xdp_act;
260 s16 offset;
261 int err;
262
263 xdp_prog = rcu_dereference_bh(efx->xdp_prog);
264 if (!xdp_prog)
265 return true;
266
267 rx_queue = efx_channel_get_rx_queue(channel);
268
269 if (unlikely(channel->rx_pkt_n_frags > 1)) {
270 /* We can't do XDP on fragmented packets - drop. */
271 efx_free_rx_buffers(rx_queue, rx_buf,
272 channel->rx_pkt_n_frags);
273 if (net_ratelimit())
274 netif_err(efx, rx_err, efx->net_dev,
275 "XDP is not possible with multiple receive fragments (%d)\n",
276 channel->rx_pkt_n_frags);
277 channel->n_rx_xdp_bad_drops++;
278 return false;
279 }
280
281 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr,
282 rx_buf->len, DMA_FROM_DEVICE);
283
284 /* Save the rx prefix. */
285 EFX_WARN_ON_PARANOID(efx->rx_prefix_size > EFX_MAX_RX_PREFIX_SIZE);
286 memcpy(rx_prefix, *ehp - efx->rx_prefix_size,
287 efx->rx_prefix_size);
288
289 xdp_init_buff(&xdp, efx->rx_page_buf_step, &rx_queue->xdp_rxq_info);
290 /* No support yet for XDP metadata */
291 xdp_prepare_buff(&xdp, *ehp - EFX_XDP_HEADROOM, EFX_XDP_HEADROOM,
292 rx_buf->len, false);
293
294 xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
295
296 offset = (u8 *)xdp.data - *ehp;
297
298 switch (xdp_act) {
299 case XDP_PASS:
300 /* Fix up rx prefix. */
301 if (offset) {
302 *ehp += offset;
303 rx_buf->page_offset += offset;
304 rx_buf->len -= offset;
305 memcpy(*ehp - efx->rx_prefix_size, rx_prefix,
306 efx->rx_prefix_size);
307 }
308 break;
309
310 case XDP_TX:
311 /* Buffer ownership passes to tx on success. */
312 xdpf = xdp_convert_buff_to_frame(&xdp);
313 err = efx_xdp_tx_buffers(efx, 1, &xdpf, true);
314 if (unlikely(err != 1)) {
315 efx_free_rx_buffers(rx_queue, rx_buf, 1);
316 if (net_ratelimit())
317 netif_err(efx, rx_err, efx->net_dev,
318 "XDP TX failed (%d)\n", err);
319 channel->n_rx_xdp_bad_drops++;
320 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
321 } else {
322 channel->n_rx_xdp_tx++;
323 }
324 break;
325
326 case XDP_REDIRECT:
327 err = xdp_do_redirect(efx->net_dev, &xdp, xdp_prog);
328 if (unlikely(err)) {
329 efx_free_rx_buffers(rx_queue, rx_buf, 1);
330 if (net_ratelimit())
331 netif_err(efx, rx_err, efx->net_dev,
332 "XDP redirect failed (%d)\n", err);
333 channel->n_rx_xdp_bad_drops++;
334 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
335 } else {
336 channel->n_rx_xdp_redirect++;
337 }
338 break;
339
340 default:
341 bpf_warn_invalid_xdp_action(efx->net_dev, xdp_prog, xdp_act);
342 efx_free_rx_buffers(rx_queue, rx_buf, 1);
343 channel->n_rx_xdp_bad_drops++;
344 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
345 break;
346
347 case XDP_ABORTED:
348 trace_xdp_exception(efx->net_dev, xdp_prog, xdp_act);
349 fallthrough;
350 case XDP_DROP:
351 efx_free_rx_buffers(rx_queue, rx_buf, 1);
352 channel->n_rx_xdp_drops++;
353 break;
354 }
355
356 return xdp_act == XDP_PASS;
357}
358
359/* Handle a received packet. Second half: Touches packet payload. */
360void __efx_rx_packet(struct efx_channel *channel)
361{
362 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
363 struct efx_nic *efx = channel->efx;
364 struct efx_rx_buffer *rx_buf =
365 efx_rx_buffer(rx_queue, channel->rx_pkt_index);
366 u8 *eh = efx_rx_buf_va(rx_buf);
367
368 /* Read length from the prefix if necessary. This already
369 * excludes the length of the prefix itself.
370 */
371 if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN) {
372 rx_buf->len = le16_to_cpup((__le16 *)
373 (eh + efx->rx_packet_len_offset));
374 /* A known issue may prevent this being filled in;
375 * if that happens, just drop the packet.
376 * Must do that in the driver since passing a zero-length
377 * packet up to the stack may cause a crash.
378 */
379 if (unlikely(!rx_buf->len)) {
380 efx_free_rx_buffers(rx_queue, rx_buf,
381 channel->rx_pkt_n_frags);
382 channel->n_rx_frm_trunc++;
383 goto out;
384 }
385 }
386
387 /* If we're in loopback test, then pass the packet directly to the
388 * loopback layer, and free the rx_buf here
389 */
390 if (unlikely(efx->loopback_selftest)) {
391 efx_loopback_rx_packet(efx, eh, rx_buf->len);
392 efx_free_rx_buffers(rx_queue, rx_buf,
393 channel->rx_pkt_n_frags);
394 goto out;
395 }
396
397 if (!efx_do_xdp(efx, channel, rx_buf, &eh))
398 goto out;
399
400 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
401 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
402
403 if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
404 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh, 0);
405 else
406 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
407out:
408 channel->rx_pkt_n_frags = 0;
409}
1/****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2011 Solarflare Communications Inc.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
9 */
10
11#include <linux/socket.h>
12#include <linux/in.h>
13#include <linux/slab.h>
14#include <linux/ip.h>
15#include <linux/tcp.h>
16#include <linux/udp.h>
17#include <linux/prefetch.h>
18#include <linux/moduleparam.h>
19#include <net/ip.h>
20#include <net/checksum.h>
21#include "net_driver.h"
22#include "efx.h"
23#include "nic.h"
24#include "selftest.h"
25#include "workarounds.h"
26
27/* Number of RX descriptors pushed at once. */
28#define EFX_RX_BATCH 8
29
30/* Maximum size of a buffer sharing a page */
31#define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
32
33/* Size of buffer allocated for skb header area. */
34#define EFX_SKB_HEADERS 64u
35
36/*
37 * rx_alloc_method - RX buffer allocation method
38 *
39 * This driver supports two methods for allocating and using RX buffers:
40 * each RX buffer may be backed by an skb or by an order-n page.
41 *
42 * When GRO is in use then the second method has a lower overhead,
43 * since we don't have to allocate then free skbs on reassembled frames.
44 *
45 * Values:
46 * - RX_ALLOC_METHOD_AUTO = 0
47 * - RX_ALLOC_METHOD_SKB = 1
48 * - RX_ALLOC_METHOD_PAGE = 2
49 *
50 * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
51 * controlled by the parameters below.
52 *
53 * - Since pushing and popping descriptors are separated by the rx_queue
54 * size, so the watermarks should be ~rxd_size.
55 * - The performance win by using page-based allocation for GRO is less
56 * than the performance hit of using page-based allocation of non-GRO,
57 * so the watermarks should reflect this.
58 *
59 * Per channel we maintain a single variable, updated by each channel:
60 *
61 * rx_alloc_level += (gro_performed ? RX_ALLOC_FACTOR_GRO :
62 * RX_ALLOC_FACTOR_SKB)
63 * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
64 * limits the hysteresis), and update the allocation strategy:
65 *
66 * rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_GRO ?
67 * RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
68 */
69static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
70
71#define RX_ALLOC_LEVEL_GRO 0x2000
72#define RX_ALLOC_LEVEL_MAX 0x3000
73#define RX_ALLOC_FACTOR_GRO 1
74#define RX_ALLOC_FACTOR_SKB (-2)
75
76/* This is the percentage fill level below which new RX descriptors
77 * will be added to the RX descriptor ring.
78 */
79static unsigned int rx_refill_threshold;
80
81/*
82 * RX maximum head room required.
83 *
84 * This must be at least 1 to prevent overflow and at least 2 to allow
85 * pipelined receives.
86 */
87#define EFX_RXD_HEAD_ROOM 2
88
89/* Offset of ethernet header within page */
90static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
91 struct efx_rx_buffer *buf)
92{
93 /* Offset is always within one page, so we don't need to consider
94 * the page order.
95 */
96 return ((unsigned int) buf->dma_addr & (PAGE_SIZE - 1)) +
97 efx->type->rx_buffer_hash_size;
98}
99static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
100{
101 return PAGE_SIZE << efx->rx_buffer_order;
102}
103
104static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
105{
106 if (buf->flags & EFX_RX_BUF_PAGE)
107 return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
108 else
109 return (u8 *)buf->u.skb->data + efx->type->rx_buffer_hash_size;
110}
111
112static inline u32 efx_rx_buf_hash(const u8 *eh)
113{
114 /* The ethernet header is always directly after any hash. */
115#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
116 return __le32_to_cpup((const __le32 *)(eh - 4));
117#else
118 const u8 *data = eh - 4;
119 return (u32)data[0] |
120 (u32)data[1] << 8 |
121 (u32)data[2] << 16 |
122 (u32)data[3] << 24;
123#endif
124}
125
126/**
127 * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
128 *
129 * @rx_queue: Efx RX queue
130 *
131 * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
132 * struct efx_rx_buffer for each one. Return a negative error code or 0
133 * on success. May fail having only inserted fewer than EFX_RX_BATCH
134 * buffers.
135 */
136static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
137{
138 struct efx_nic *efx = rx_queue->efx;
139 struct net_device *net_dev = efx->net_dev;
140 struct efx_rx_buffer *rx_buf;
141 struct sk_buff *skb;
142 int skb_len = efx->rx_buffer_len;
143 unsigned index, count;
144
145 for (count = 0; count < EFX_RX_BATCH; ++count) {
146 index = rx_queue->added_count & rx_queue->ptr_mask;
147 rx_buf = efx_rx_buffer(rx_queue, index);
148
149 rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
150 if (unlikely(!skb))
151 return -ENOMEM;
152
153 /* Adjust the SKB for padding */
154 skb_reserve(skb, NET_IP_ALIGN);
155 rx_buf->len = skb_len - NET_IP_ALIGN;
156 rx_buf->flags = 0;
157
158 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
159 skb->data, rx_buf->len,
160 PCI_DMA_FROMDEVICE);
161 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
162 rx_buf->dma_addr))) {
163 dev_kfree_skb_any(skb);
164 rx_buf->u.skb = NULL;
165 return -EIO;
166 }
167
168 ++rx_queue->added_count;
169 ++rx_queue->alloc_skb_count;
170 }
171
172 return 0;
173}
174
175/**
176 * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
177 *
178 * @rx_queue: Efx RX queue
179 *
180 * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
181 * and populates struct efx_rx_buffers for each one. Return a negative error
182 * code or 0 on success. If a single page can be split between two buffers,
183 * then the page will either be inserted fully, or not at at all.
184 */
185static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
186{
187 struct efx_nic *efx = rx_queue->efx;
188 struct efx_rx_buffer *rx_buf;
189 struct page *page;
190 void *page_addr;
191 struct efx_rx_page_state *state;
192 dma_addr_t dma_addr;
193 unsigned index, count;
194
195 /* We can split a page between two buffers */
196 BUILD_BUG_ON(EFX_RX_BATCH & 1);
197
198 for (count = 0; count < EFX_RX_BATCH; ++count) {
199 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
200 efx->rx_buffer_order);
201 if (unlikely(page == NULL))
202 return -ENOMEM;
203 dma_addr = pci_map_page(efx->pci_dev, page, 0,
204 efx_rx_buf_size(efx),
205 PCI_DMA_FROMDEVICE);
206 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
207 __free_pages(page, efx->rx_buffer_order);
208 return -EIO;
209 }
210 page_addr = page_address(page);
211 state = page_addr;
212 state->refcnt = 0;
213 state->dma_addr = dma_addr;
214
215 page_addr += sizeof(struct efx_rx_page_state);
216 dma_addr += sizeof(struct efx_rx_page_state);
217
218 split:
219 index = rx_queue->added_count & rx_queue->ptr_mask;
220 rx_buf = efx_rx_buffer(rx_queue, index);
221 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
222 rx_buf->u.page = page;
223 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
224 rx_buf->flags = EFX_RX_BUF_PAGE;
225 ++rx_queue->added_count;
226 ++rx_queue->alloc_page_count;
227 ++state->refcnt;
228
229 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
230 /* Use the second half of the page */
231 get_page(page);
232 dma_addr += (PAGE_SIZE >> 1);
233 page_addr += (PAGE_SIZE >> 1);
234 ++count;
235 goto split;
236 }
237 }
238
239 return 0;
240}
241
242static void efx_unmap_rx_buffer(struct efx_nic *efx,
243 struct efx_rx_buffer *rx_buf)
244{
245 if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
246 struct efx_rx_page_state *state;
247
248 state = page_address(rx_buf->u.page);
249 if (--state->refcnt == 0) {
250 pci_unmap_page(efx->pci_dev,
251 state->dma_addr,
252 efx_rx_buf_size(efx),
253 PCI_DMA_FROMDEVICE);
254 }
255 } else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
256 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
257 rx_buf->len, PCI_DMA_FROMDEVICE);
258 }
259}
260
261static void efx_free_rx_buffer(struct efx_nic *efx,
262 struct efx_rx_buffer *rx_buf)
263{
264 if ((rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.page) {
265 __free_pages(rx_buf->u.page, efx->rx_buffer_order);
266 rx_buf->u.page = NULL;
267 } else if (!(rx_buf->flags & EFX_RX_BUF_PAGE) && rx_buf->u.skb) {
268 dev_kfree_skb_any(rx_buf->u.skb);
269 rx_buf->u.skb = NULL;
270 }
271}
272
273static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
274 struct efx_rx_buffer *rx_buf)
275{
276 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
277 efx_free_rx_buffer(rx_queue->efx, rx_buf);
278}
279
280/* Attempt to resurrect the other receive buffer that used to share this page,
281 * which had previously been passed up to the kernel and freed. */
282static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
283 struct efx_rx_buffer *rx_buf)
284{
285 struct efx_rx_page_state *state = page_address(rx_buf->u.page);
286 struct efx_rx_buffer *new_buf;
287 unsigned fill_level, index;
288
289 /* +1 because efx_rx_packet() incremented removed_count. +1 because
290 * we'd like to insert an additional descriptor whilst leaving
291 * EFX_RXD_HEAD_ROOM for the non-recycle path */
292 fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
293 if (unlikely(fill_level > rx_queue->max_fill)) {
294 /* We could place "state" on a list, and drain the list in
295 * efx_fast_push_rx_descriptors(). For now, this will do. */
296 return;
297 }
298
299 ++state->refcnt;
300 get_page(rx_buf->u.page);
301
302 index = rx_queue->added_count & rx_queue->ptr_mask;
303 new_buf = efx_rx_buffer(rx_queue, index);
304 new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
305 new_buf->u.page = rx_buf->u.page;
306 new_buf->len = rx_buf->len;
307 new_buf->flags = EFX_RX_BUF_PAGE;
308 ++rx_queue->added_count;
309}
310
311/* Recycle the given rx buffer directly back into the rx_queue. There is
312 * always room to add this buffer, because we've just popped a buffer. */
313static void efx_recycle_rx_buffer(struct efx_channel *channel,
314 struct efx_rx_buffer *rx_buf)
315{
316 struct efx_nic *efx = channel->efx;
317 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
318 struct efx_rx_buffer *new_buf;
319 unsigned index;
320
321 rx_buf->flags &= EFX_RX_BUF_PAGE;
322
323 if ((rx_buf->flags & EFX_RX_BUF_PAGE) &&
324 efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
325 page_count(rx_buf->u.page) == 1)
326 efx_resurrect_rx_buffer(rx_queue, rx_buf);
327
328 index = rx_queue->added_count & rx_queue->ptr_mask;
329 new_buf = efx_rx_buffer(rx_queue, index);
330
331 memcpy(new_buf, rx_buf, sizeof(*new_buf));
332 rx_buf->u.page = NULL;
333 ++rx_queue->added_count;
334}
335
336/**
337 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
338 * @rx_queue: RX descriptor queue
339 * This will aim to fill the RX descriptor queue up to
340 * @rx_queue->@max_fill. If there is insufficient atomic
341 * memory to do so, a slow fill will be scheduled.
342 *
343 * The caller must provide serialisation (none is used here). In practise,
344 * this means this function must run from the NAPI handler, or be called
345 * when NAPI is disabled.
346 */
347void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
348{
349 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
350 unsigned fill_level;
351 int space, rc = 0;
352
353 /* Calculate current fill level, and exit if we don't need to fill */
354 fill_level = (rx_queue->added_count - rx_queue->removed_count);
355 EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
356 if (fill_level >= rx_queue->fast_fill_trigger)
357 goto out;
358
359 /* Record minimum fill level */
360 if (unlikely(fill_level < rx_queue->min_fill)) {
361 if (fill_level)
362 rx_queue->min_fill = fill_level;
363 }
364
365 space = rx_queue->max_fill - fill_level;
366 EFX_BUG_ON_PARANOID(space < EFX_RX_BATCH);
367
368 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
369 "RX queue %d fast-filling descriptor ring from"
370 " level %d to level %d using %s allocation\n",
371 efx_rx_queue_index(rx_queue), fill_level,
372 rx_queue->max_fill,
373 channel->rx_alloc_push_pages ? "page" : "skb");
374
375 do {
376 if (channel->rx_alloc_push_pages)
377 rc = efx_init_rx_buffers_page(rx_queue);
378 else
379 rc = efx_init_rx_buffers_skb(rx_queue);
380 if (unlikely(rc)) {
381 /* Ensure that we don't leave the rx queue empty */
382 if (rx_queue->added_count == rx_queue->removed_count)
383 efx_schedule_slow_fill(rx_queue);
384 goto out;
385 }
386 } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
387
388 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
389 "RX queue %d fast-filled descriptor ring "
390 "to level %d\n", efx_rx_queue_index(rx_queue),
391 rx_queue->added_count - rx_queue->removed_count);
392
393 out:
394 if (rx_queue->notified_count != rx_queue->added_count)
395 efx_nic_notify_rx_desc(rx_queue);
396}
397
398void efx_rx_slow_fill(unsigned long context)
399{
400 struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
401
402 /* Post an event to cause NAPI to run and refill the queue */
403 efx_nic_generate_fill_event(rx_queue);
404 ++rx_queue->slow_fill_count;
405}
406
407static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
408 struct efx_rx_buffer *rx_buf,
409 int len, bool *leak_packet)
410{
411 struct efx_nic *efx = rx_queue->efx;
412 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
413
414 if (likely(len <= max_len))
415 return;
416
417 /* The packet must be discarded, but this is only a fatal error
418 * if the caller indicated it was
419 */
420 rx_buf->flags |= EFX_RX_PKT_DISCARD;
421
422 if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
423 if (net_ratelimit())
424 netif_err(efx, rx_err, efx->net_dev,
425 " RX queue %d seriously overlength "
426 "RX event (0x%x > 0x%x+0x%x). Leaking\n",
427 efx_rx_queue_index(rx_queue), len, max_len,
428 efx->type->rx_buffer_padding);
429 /* If this buffer was skb-allocated, then the meta
430 * data at the end of the skb will be trashed. So
431 * we have no choice but to leak the fragment.
432 */
433 *leak_packet = !(rx_buf->flags & EFX_RX_BUF_PAGE);
434 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
435 } else {
436 if (net_ratelimit())
437 netif_err(efx, rx_err, efx->net_dev,
438 " RX queue %d overlength RX event "
439 "(0x%x > 0x%x)\n",
440 efx_rx_queue_index(rx_queue), len, max_len);
441 }
442
443 efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
444}
445
446/* Pass a received packet up through GRO. GRO can handle pages
447 * regardless of checksum state and skbs with a good checksum.
448 */
449static void efx_rx_packet_gro(struct efx_channel *channel,
450 struct efx_rx_buffer *rx_buf,
451 const u8 *eh)
452{
453 struct napi_struct *napi = &channel->napi_str;
454 gro_result_t gro_result;
455
456 if (rx_buf->flags & EFX_RX_BUF_PAGE) {
457 struct efx_nic *efx = channel->efx;
458 struct page *page = rx_buf->u.page;
459 struct sk_buff *skb;
460
461 rx_buf->u.page = NULL;
462
463 skb = napi_get_frags(napi);
464 if (!skb) {
465 put_page(page);
466 return;
467 }
468
469 if (efx->net_dev->features & NETIF_F_RXHASH)
470 skb->rxhash = efx_rx_buf_hash(eh);
471
472 skb_fill_page_desc(skb, 0, page,
473 efx_rx_buf_offset(efx, rx_buf), rx_buf->len);
474
475 skb->len = rx_buf->len;
476 skb->data_len = rx_buf->len;
477 skb->truesize += rx_buf->len;
478 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
479 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
480
481 skb_record_rx_queue(skb, channel->channel);
482
483 gro_result = napi_gro_frags(napi);
484 } else {
485 struct sk_buff *skb = rx_buf->u.skb;
486
487 EFX_BUG_ON_PARANOID(!(rx_buf->flags & EFX_RX_PKT_CSUMMED));
488 rx_buf->u.skb = NULL;
489 skb->ip_summed = CHECKSUM_UNNECESSARY;
490
491 gro_result = napi_gro_receive(napi, skb);
492 }
493
494 if (gro_result == GRO_NORMAL) {
495 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
496 } else if (gro_result != GRO_DROP) {
497 channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
498 channel->irq_mod_score += 2;
499 }
500}
501
502void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
503 unsigned int len, u16 flags)
504{
505 struct efx_nic *efx = rx_queue->efx;
506 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
507 struct efx_rx_buffer *rx_buf;
508 bool leak_packet = false;
509
510 rx_buf = efx_rx_buffer(rx_queue, index);
511 rx_buf->flags |= flags;
512
513 /* This allows the refill path to post another buffer.
514 * EFX_RXD_HEAD_ROOM ensures that the slot we are using
515 * isn't overwritten yet.
516 */
517 rx_queue->removed_count++;
518
519 /* Validate the length encoded in the event vs the descriptor pushed */
520 efx_rx_packet__check_len(rx_queue, rx_buf, len, &leak_packet);
521
522 netif_vdbg(efx, rx_status, efx->net_dev,
523 "RX queue %d received id %x at %llx+%x %s%s\n",
524 efx_rx_queue_index(rx_queue), index,
525 (unsigned long long)rx_buf->dma_addr, len,
526 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
527 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
528
529 /* Discard packet, if instructed to do so */
530 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
531 if (unlikely(leak_packet))
532 channel->n_skbuff_leaks++;
533 else
534 efx_recycle_rx_buffer(channel, rx_buf);
535
536 /* Don't hold off the previous receive */
537 rx_buf = NULL;
538 goto out;
539 }
540
541 /* Release card resources - assumes all RX buffers consumed in-order
542 * per RX queue
543 */
544 efx_unmap_rx_buffer(efx, rx_buf);
545
546 /* Prefetch nice and early so data will (hopefully) be in cache by
547 * the time we look at it.
548 */
549 prefetch(efx_rx_buf_eh(efx, rx_buf));
550
551 /* Pipeline receives so that we give time for packet headers to be
552 * prefetched into cache.
553 */
554 rx_buf->len = len - efx->type->rx_buffer_hash_size;
555out:
556 if (channel->rx_pkt)
557 __efx_rx_packet(channel, channel->rx_pkt);
558 channel->rx_pkt = rx_buf;
559}
560
561static void efx_rx_deliver(struct efx_channel *channel,
562 struct efx_rx_buffer *rx_buf)
563{
564 struct sk_buff *skb;
565
566 /* We now own the SKB */
567 skb = rx_buf->u.skb;
568 rx_buf->u.skb = NULL;
569
570 /* Set the SKB flags */
571 skb_checksum_none_assert(skb);
572
573 /* Pass the packet up */
574 netif_receive_skb(skb);
575
576 /* Update allocation strategy method */
577 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
578}
579
580/* Handle a received packet. Second half: Touches packet payload. */
581void __efx_rx_packet(struct efx_channel *channel, struct efx_rx_buffer *rx_buf)
582{
583 struct efx_nic *efx = channel->efx;
584 u8 *eh = efx_rx_buf_eh(efx, rx_buf);
585
586 /* If we're in loopback test, then pass the packet directly to the
587 * loopback layer, and free the rx_buf here
588 */
589 if (unlikely(efx->loopback_selftest)) {
590 efx_loopback_rx_packet(efx, eh, rx_buf->len);
591 efx_free_rx_buffer(efx, rx_buf);
592 return;
593 }
594
595 if (!(rx_buf->flags & EFX_RX_BUF_PAGE)) {
596 struct sk_buff *skb = rx_buf->u.skb;
597
598 prefetch(skb_shinfo(skb));
599
600 skb_reserve(skb, efx->type->rx_buffer_hash_size);
601 skb_put(skb, rx_buf->len);
602
603 if (efx->net_dev->features & NETIF_F_RXHASH)
604 skb->rxhash = efx_rx_buf_hash(eh);
605
606 /* Move past the ethernet header. rx_buf->data still points
607 * at the ethernet header */
608 skb->protocol = eth_type_trans(skb, efx->net_dev);
609
610 skb_record_rx_queue(skb, channel->channel);
611 }
612
613 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
614 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
615
616 if (likely(rx_buf->flags & (EFX_RX_BUF_PAGE | EFX_RX_PKT_CSUMMED)))
617 efx_rx_packet_gro(channel, rx_buf, eh);
618 else
619 efx_rx_deliver(channel, rx_buf);
620}
621
622void efx_rx_strategy(struct efx_channel *channel)
623{
624 enum efx_rx_alloc_method method = rx_alloc_method;
625
626 /* Only makes sense to use page based allocation if GRO is enabled */
627 if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
628 method = RX_ALLOC_METHOD_SKB;
629 } else if (method == RX_ALLOC_METHOD_AUTO) {
630 /* Constrain the rx_alloc_level */
631 if (channel->rx_alloc_level < 0)
632 channel->rx_alloc_level = 0;
633 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
634 channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
635
636 /* Decide on the allocation method */
637 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
638 RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
639 }
640
641 /* Push the option */
642 channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
643}
644
645int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
646{
647 struct efx_nic *efx = rx_queue->efx;
648 unsigned int entries;
649 int rc;
650
651 /* Create the smallest power-of-two aligned ring */
652 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
653 EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
654 rx_queue->ptr_mask = entries - 1;
655
656 netif_dbg(efx, probe, efx->net_dev,
657 "creating RX queue %d size %#x mask %#x\n",
658 efx_rx_queue_index(rx_queue), efx->rxq_entries,
659 rx_queue->ptr_mask);
660
661 /* Allocate RX buffers */
662 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
663 GFP_KERNEL);
664 if (!rx_queue->buffer)
665 return -ENOMEM;
666
667 rc = efx_nic_probe_rx(rx_queue);
668 if (rc) {
669 kfree(rx_queue->buffer);
670 rx_queue->buffer = NULL;
671 }
672 return rc;
673}
674
675void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
676{
677 struct efx_nic *efx = rx_queue->efx;
678 unsigned int max_fill, trigger, max_trigger;
679
680 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
681 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
682
683 /* Initialise ptr fields */
684 rx_queue->added_count = 0;
685 rx_queue->notified_count = 0;
686 rx_queue->removed_count = 0;
687 rx_queue->min_fill = -1U;
688
689 /* Initialise limit fields */
690 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
691 max_trigger = max_fill - EFX_RX_BATCH;
692 if (rx_refill_threshold != 0) {
693 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
694 if (trigger > max_trigger)
695 trigger = max_trigger;
696 } else {
697 trigger = max_trigger;
698 }
699
700 rx_queue->max_fill = max_fill;
701 rx_queue->fast_fill_trigger = trigger;
702
703 /* Set up RX descriptor ring */
704 rx_queue->enabled = true;
705 efx_nic_init_rx(rx_queue);
706}
707
708void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
709{
710 int i;
711 struct efx_rx_buffer *rx_buf;
712
713 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
714 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
715
716 /* A flush failure might have left rx_queue->enabled */
717 rx_queue->enabled = false;
718
719 del_timer_sync(&rx_queue->slow_fill);
720 efx_nic_fini_rx(rx_queue);
721
722 /* Release RX buffers NB start at index 0 not current HW ptr */
723 if (rx_queue->buffer) {
724 for (i = 0; i <= rx_queue->ptr_mask; i++) {
725 rx_buf = efx_rx_buffer(rx_queue, i);
726 efx_fini_rx_buffer(rx_queue, rx_buf);
727 }
728 }
729}
730
731void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
732{
733 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
734 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
735
736 efx_nic_remove_rx(rx_queue);
737
738 kfree(rx_queue->buffer);
739 rx_queue->buffer = NULL;
740}
741
742
743module_param(rx_alloc_method, int, 0644);
744MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
745
746module_param(rx_refill_threshold, uint, 0444);
747MODULE_PARM_DESC(rx_refill_threshold,
748 "RX descriptor ring refill threshold (%)");
749