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// 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_driver.h"
21#include "efx.h"
22#include "filter.h"
23#include "nic.h"
24#include "selftest.h"
25#include "workarounds.h"
26
27/* Preferred number of descriptors to fill at once */
28#define EFX_RX_PREFERRED_BATCH 8U
29
30/* Number of RX buffers to recycle pages for. When creating the RX page recycle
31 * ring, this number is divided by the number of buffers per page to calculate
32 * the number of pages to store in the RX page recycle ring.
33 */
34#define EFX_RECYCLE_RING_SIZE_IOMMU 4096
35#define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
36
37/* Size of buffer allocated for skb header area. */
38#define EFX_SKB_HEADERS 128u
39
40/* This is the percentage fill level below which new RX descriptors
41 * will be added to the RX descriptor ring.
42 */
43static unsigned int rx_refill_threshold;
44
45/* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
46#define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
47 EFX_RX_USR_BUF_SIZE)
48
49/*
50 * RX maximum head room required.
51 *
52 * This must be at least 1 to prevent overflow, plus one packet-worth
53 * to allow pipelined receives.
54 */
55#define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
56
57static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
58{
59 return page_address(buf->page) + buf->page_offset;
60}
61
62static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
63{
64#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
65 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
66#else
67 const u8 *data = eh + efx->rx_packet_hash_offset;
68 return (u32)data[0] |
69 (u32)data[1] << 8 |
70 (u32)data[2] << 16 |
71 (u32)data[3] << 24;
72#endif
73}
74
75static inline struct efx_rx_buffer *
76efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
77{
78 if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
79 return efx_rx_buffer(rx_queue, 0);
80 else
81 return rx_buf + 1;
82}
83
84static inline void efx_sync_rx_buffer(struct efx_nic *efx,
85 struct efx_rx_buffer *rx_buf,
86 unsigned int len)
87{
88 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
89 DMA_FROM_DEVICE);
90}
91
92void efx_rx_config_page_split(struct efx_nic *efx)
93{
94 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
95 EFX_RX_BUF_ALIGNMENT);
96 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
97 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
98 efx->rx_page_buf_step);
99 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
100 efx->rx_bufs_per_page;
101 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
102 efx->rx_bufs_per_page);
103}
104
105/* Check the RX page recycle ring for a page that can be reused. */
106static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
107{
108 struct efx_nic *efx = rx_queue->efx;
109 struct page *page;
110 struct efx_rx_page_state *state;
111 unsigned index;
112
113 index = rx_queue->page_remove & rx_queue->page_ptr_mask;
114 page = rx_queue->page_ring[index];
115 if (page == NULL)
116 return NULL;
117
118 rx_queue->page_ring[index] = NULL;
119 /* page_remove cannot exceed page_add. */
120 if (rx_queue->page_remove != rx_queue->page_add)
121 ++rx_queue->page_remove;
122
123 /* If page_count is 1 then we hold the only reference to this page. */
124 if (page_count(page) == 1) {
125 ++rx_queue->page_recycle_count;
126 return page;
127 } else {
128 state = page_address(page);
129 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
130 PAGE_SIZE << efx->rx_buffer_order,
131 DMA_FROM_DEVICE);
132 put_page(page);
133 ++rx_queue->page_recycle_failed;
134 }
135
136 return NULL;
137}
138
139/**
140 * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
141 *
142 * @rx_queue: Efx RX queue
143 *
144 * This allocates a batch of pages, maps them for DMA, and populates
145 * struct efx_rx_buffers for each one. Return a negative error code or
146 * 0 on success. If a single page can be used for multiple buffers,
147 * then the page will either be inserted fully, or not at all.
148 */
149static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
150{
151 struct efx_nic *efx = rx_queue->efx;
152 struct efx_rx_buffer *rx_buf;
153 struct page *page;
154 unsigned int page_offset;
155 struct efx_rx_page_state *state;
156 dma_addr_t dma_addr;
157 unsigned index, count;
158
159 count = 0;
160 do {
161 page = efx_reuse_page(rx_queue);
162 if (page == NULL) {
163 page = alloc_pages(__GFP_COMP |
164 (atomic ? GFP_ATOMIC : GFP_KERNEL),
165 efx->rx_buffer_order);
166 if (unlikely(page == NULL))
167 return -ENOMEM;
168 dma_addr =
169 dma_map_page(&efx->pci_dev->dev, page, 0,
170 PAGE_SIZE << efx->rx_buffer_order,
171 DMA_FROM_DEVICE);
172 if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
173 dma_addr))) {
174 __free_pages(page, efx->rx_buffer_order);
175 return -EIO;
176 }
177 state = page_address(page);
178 state->dma_addr = dma_addr;
179 } else {
180 state = page_address(page);
181 dma_addr = state->dma_addr;
182 }
183
184 dma_addr += sizeof(struct efx_rx_page_state);
185 page_offset = sizeof(struct efx_rx_page_state);
186
187 do {
188 index = rx_queue->added_count & rx_queue->ptr_mask;
189 rx_buf = efx_rx_buffer(rx_queue, index);
190 rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
191 rx_buf->page = page;
192 rx_buf->page_offset = page_offset + efx->rx_ip_align;
193 rx_buf->len = efx->rx_dma_len;
194 rx_buf->flags = 0;
195 ++rx_queue->added_count;
196 get_page(page);
197 dma_addr += efx->rx_page_buf_step;
198 page_offset += efx->rx_page_buf_step;
199 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
200
201 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
202 } while (++count < efx->rx_pages_per_batch);
203
204 return 0;
205}
206
207/* Unmap a DMA-mapped page. This function is only called for the final RX
208 * buffer in a page.
209 */
210static void efx_unmap_rx_buffer(struct efx_nic *efx,
211 struct efx_rx_buffer *rx_buf)
212{
213 struct page *page = rx_buf->page;
214
215 if (page) {
216 struct efx_rx_page_state *state = page_address(page);
217 dma_unmap_page(&efx->pci_dev->dev,
218 state->dma_addr,
219 PAGE_SIZE << efx->rx_buffer_order,
220 DMA_FROM_DEVICE);
221 }
222}
223
224static void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
225 struct efx_rx_buffer *rx_buf,
226 unsigned int num_bufs)
227{
228 do {
229 if (rx_buf->page) {
230 put_page(rx_buf->page);
231 rx_buf->page = NULL;
232 }
233 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
234 } while (--num_bufs);
235}
236
237/* Attempt to recycle the page if there is an RX recycle ring; the page can
238 * only be added if this is the final RX buffer, to prevent pages being used in
239 * the descriptor ring and appearing in the recycle ring simultaneously.
240 */
241static void efx_recycle_rx_page(struct efx_channel *channel,
242 struct efx_rx_buffer *rx_buf)
243{
244 struct page *page = rx_buf->page;
245 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
246 struct efx_nic *efx = rx_queue->efx;
247 unsigned index;
248
249 /* Only recycle the page after processing the final buffer. */
250 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
251 return;
252
253 index = rx_queue->page_add & rx_queue->page_ptr_mask;
254 if (rx_queue->page_ring[index] == NULL) {
255 unsigned read_index = rx_queue->page_remove &
256 rx_queue->page_ptr_mask;
257
258 /* The next slot in the recycle ring is available, but
259 * increment page_remove if the read pointer currently
260 * points here.
261 */
262 if (read_index == index)
263 ++rx_queue->page_remove;
264 rx_queue->page_ring[index] = page;
265 ++rx_queue->page_add;
266 return;
267 }
268 ++rx_queue->page_recycle_full;
269 efx_unmap_rx_buffer(efx, rx_buf);
270 put_page(rx_buf->page);
271}
272
273static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
274 struct efx_rx_buffer *rx_buf)
275{
276 /* Release the page reference we hold for the buffer. */
277 if (rx_buf->page)
278 put_page(rx_buf->page);
279
280 /* If this is the last buffer in a page, unmap and free it. */
281 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
282 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
283 efx_free_rx_buffers(rx_queue, rx_buf, 1);
284 }
285 rx_buf->page = NULL;
286}
287
288/* Recycle the pages that are used by buffers that have just been received. */
289static void efx_recycle_rx_pages(struct efx_channel *channel,
290 struct efx_rx_buffer *rx_buf,
291 unsigned int n_frags)
292{
293 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
294
295 do {
296 efx_recycle_rx_page(channel, rx_buf);
297 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
298 } while (--n_frags);
299}
300
301static void efx_discard_rx_packet(struct efx_channel *channel,
302 struct efx_rx_buffer *rx_buf,
303 unsigned int n_frags)
304{
305 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
306
307 efx_recycle_rx_pages(channel, rx_buf, n_frags);
308
309 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
310}
311
312/**
313 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
314 * @rx_queue: RX descriptor queue
315 *
316 * This will aim to fill the RX descriptor queue up to
317 * @rx_queue->@max_fill. If there is insufficient atomic
318 * memory to do so, a slow fill will be scheduled.
319 *
320 * The caller must provide serialisation (none is used here). In practise,
321 * this means this function must run from the NAPI handler, or be called
322 * when NAPI is disabled.
323 */
324void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
325{
326 struct efx_nic *efx = rx_queue->efx;
327 unsigned int fill_level, batch_size;
328 int space, rc = 0;
329
330 if (!rx_queue->refill_enabled)
331 return;
332
333 /* Calculate current fill level, and exit if we don't need to fill */
334 fill_level = (rx_queue->added_count - rx_queue->removed_count);
335 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
336 if (fill_level >= rx_queue->fast_fill_trigger)
337 goto out;
338
339 /* Record minimum fill level */
340 if (unlikely(fill_level < rx_queue->min_fill)) {
341 if (fill_level)
342 rx_queue->min_fill = fill_level;
343 }
344
345 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
346 space = rx_queue->max_fill - fill_level;
347 EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
348
349 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
350 "RX queue %d fast-filling descriptor ring from"
351 " level %d to level %d\n",
352 efx_rx_queue_index(rx_queue), fill_level,
353 rx_queue->max_fill);
354
355
356 do {
357 rc = efx_init_rx_buffers(rx_queue, atomic);
358 if (unlikely(rc)) {
359 /* Ensure that we don't leave the rx queue empty */
360 efx_schedule_slow_fill(rx_queue);
361 goto out;
362 }
363 } while ((space -= batch_size) >= batch_size);
364
365 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
366 "RX queue %d fast-filled descriptor ring "
367 "to level %d\n", efx_rx_queue_index(rx_queue),
368 rx_queue->added_count - rx_queue->removed_count);
369
370 out:
371 if (rx_queue->notified_count != rx_queue->added_count)
372 efx_nic_notify_rx_desc(rx_queue);
373}
374
375void efx_rx_slow_fill(struct timer_list *t)
376{
377 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
378
379 /* Post an event to cause NAPI to run and refill the queue */
380 efx_nic_generate_fill_event(rx_queue);
381 ++rx_queue->slow_fill_count;
382}
383
384static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
385 struct efx_rx_buffer *rx_buf,
386 int len)
387{
388 struct efx_nic *efx = rx_queue->efx;
389 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
390
391 if (likely(len <= max_len))
392 return;
393
394 /* The packet must be discarded, but this is only a fatal error
395 * if the caller indicated it was
396 */
397 rx_buf->flags |= EFX_RX_PKT_DISCARD;
398
399 if (net_ratelimit())
400 netif_err(efx, rx_err, efx->net_dev,
401 "RX queue %d overlength RX event (%#x > %#x)\n",
402 efx_rx_queue_index(rx_queue), len, max_len);
403
404 efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
405}
406
407/* Pass a received packet up through GRO. GRO can handle pages
408 * regardless of checksum state and skbs with a good checksum.
409 */
410static void
411efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
412 unsigned int n_frags, u8 *eh)
413{
414 struct napi_struct *napi = &channel->napi_str;
415 struct efx_nic *efx = channel->efx;
416 struct sk_buff *skb;
417
418 skb = napi_get_frags(napi);
419 if (unlikely(!skb)) {
420 struct efx_rx_queue *rx_queue;
421
422 rx_queue = efx_channel_get_rx_queue(channel);
423 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
424 return;
425 }
426
427 if (efx->net_dev->features & NETIF_F_RXHASH)
428 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
429 PKT_HASH_TYPE_L3);
430 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
431 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
432 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
433
434 for (;;) {
435 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
436 rx_buf->page, rx_buf->page_offset,
437 rx_buf->len);
438 rx_buf->page = NULL;
439 skb->len += rx_buf->len;
440 if (skb_shinfo(skb)->nr_frags == n_frags)
441 break;
442
443 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
444 }
445
446 skb->data_len = skb->len;
447 skb->truesize += n_frags * efx->rx_buffer_truesize;
448
449 skb_record_rx_queue(skb, channel->rx_queue.core_index);
450
451 napi_gro_frags(napi);
452}
453
454/* Allocate and construct an SKB around page fragments */
455static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
456 struct efx_rx_buffer *rx_buf,
457 unsigned int n_frags,
458 u8 *eh, int hdr_len)
459{
460 struct efx_nic *efx = channel->efx;
461 struct sk_buff *skb;
462
463 /* Allocate an SKB to store the headers */
464 skb = netdev_alloc_skb(efx->net_dev,
465 efx->rx_ip_align + efx->rx_prefix_size +
466 hdr_len);
467 if (unlikely(skb == NULL)) {
468 atomic_inc(&efx->n_rx_noskb_drops);
469 return NULL;
470 }
471
472 EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
473
474 memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
475 efx->rx_prefix_size + hdr_len);
476 skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
477 __skb_put(skb, hdr_len);
478
479 /* Append the remaining page(s) onto the frag list */
480 if (rx_buf->len > hdr_len) {
481 rx_buf->page_offset += hdr_len;
482 rx_buf->len -= hdr_len;
483
484 for (;;) {
485 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
486 rx_buf->page, rx_buf->page_offset,
487 rx_buf->len);
488 rx_buf->page = NULL;
489 skb->len += rx_buf->len;
490 skb->data_len += rx_buf->len;
491 if (skb_shinfo(skb)->nr_frags == n_frags)
492 break;
493
494 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
495 }
496 } else {
497 __free_pages(rx_buf->page, efx->rx_buffer_order);
498 rx_buf->page = NULL;
499 n_frags = 0;
500 }
501
502 skb->truesize += n_frags * efx->rx_buffer_truesize;
503
504 /* Move past the ethernet header */
505 skb->protocol = eth_type_trans(skb, efx->net_dev);
506
507 skb_mark_napi_id(skb, &channel->napi_str);
508
509 return skb;
510}
511
512void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
513 unsigned int n_frags, unsigned int len, u16 flags)
514{
515 struct efx_nic *efx = rx_queue->efx;
516 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
517 struct efx_rx_buffer *rx_buf;
518
519 rx_queue->rx_packets++;
520
521 rx_buf = efx_rx_buffer(rx_queue, index);
522 rx_buf->flags |= flags;
523
524 /* Validate the number of fragments and completed length */
525 if (n_frags == 1) {
526 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
527 efx_rx_packet__check_len(rx_queue, rx_buf, len);
528 } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
529 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
530 unlikely(len > n_frags * efx->rx_dma_len) ||
531 unlikely(!efx->rx_scatter)) {
532 /* If this isn't an explicit discard request, either
533 * the hardware or the driver is broken.
534 */
535 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
536 rx_buf->flags |= EFX_RX_PKT_DISCARD;
537 }
538
539 netif_vdbg(efx, rx_status, efx->net_dev,
540 "RX queue %d received ids %x-%x len %d %s%s\n",
541 efx_rx_queue_index(rx_queue), index,
542 (index + n_frags - 1) & rx_queue->ptr_mask, len,
543 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
544 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
545
546 /* Discard packet, if instructed to do so. Process the
547 * previous receive first.
548 */
549 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
550 efx_rx_flush_packet(channel);
551 efx_discard_rx_packet(channel, rx_buf, n_frags);
552 return;
553 }
554
555 if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
556 rx_buf->len = len;
557
558 /* Release and/or sync the DMA mapping - assumes all RX buffers
559 * consumed in-order per RX queue.
560 */
561 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
562
563 /* Prefetch nice and early so data will (hopefully) be in cache by
564 * the time we look at it.
565 */
566 prefetch(efx_rx_buf_va(rx_buf));
567
568 rx_buf->page_offset += efx->rx_prefix_size;
569 rx_buf->len -= efx->rx_prefix_size;
570
571 if (n_frags > 1) {
572 /* Release/sync DMA mapping for additional fragments.
573 * Fix length for last fragment.
574 */
575 unsigned int tail_frags = n_frags - 1;
576
577 for (;;) {
578 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
579 if (--tail_frags == 0)
580 break;
581 efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
582 }
583 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
584 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
585 }
586
587 /* All fragments have been DMA-synced, so recycle pages. */
588 rx_buf = efx_rx_buffer(rx_queue, index);
589 efx_recycle_rx_pages(channel, rx_buf, n_frags);
590
591 /* Pipeline receives so that we give time for packet headers to be
592 * prefetched into cache.
593 */
594 efx_rx_flush_packet(channel);
595 channel->rx_pkt_n_frags = n_frags;
596 channel->rx_pkt_index = index;
597}
598
599static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
600 struct efx_rx_buffer *rx_buf,
601 unsigned int n_frags)
602{
603 struct sk_buff *skb;
604 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
605
606 skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
607 if (unlikely(skb == NULL)) {
608 struct efx_rx_queue *rx_queue;
609
610 rx_queue = efx_channel_get_rx_queue(channel);
611 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
612 return;
613 }
614 skb_record_rx_queue(skb, channel->rx_queue.core_index);
615
616 /* Set the SKB flags */
617 skb_checksum_none_assert(skb);
618 if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
619 skb->ip_summed = CHECKSUM_UNNECESSARY;
620 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
621 }
622
623 efx_rx_skb_attach_timestamp(channel, skb);
624
625 if (channel->type->receive_skb)
626 if (channel->type->receive_skb(channel, skb))
627 return;
628
629 /* Pass the packet up */
630 if (channel->rx_list != NULL)
631 /* Add to list, will pass up later */
632 list_add_tail(&skb->list, channel->rx_list);
633 else
634 /* No list, so pass it up now */
635 netif_receive_skb(skb);
636}
637
638/* Handle a received packet. Second half: Touches packet payload. */
639void __efx_rx_packet(struct efx_channel *channel)
640{
641 struct efx_nic *efx = channel->efx;
642 struct efx_rx_buffer *rx_buf =
643 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
644 u8 *eh = efx_rx_buf_va(rx_buf);
645
646 /* Read length from the prefix if necessary. This already
647 * excludes the length of the prefix itself.
648 */
649 if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
650 rx_buf->len = le16_to_cpup((__le16 *)
651 (eh + efx->rx_packet_len_offset));
652
653 /* If we're in loopback test, then pass the packet directly to the
654 * loopback layer, and free the rx_buf here
655 */
656 if (unlikely(efx->loopback_selftest)) {
657 struct efx_rx_queue *rx_queue;
658
659 efx_loopback_rx_packet(efx, eh, rx_buf->len);
660 rx_queue = efx_channel_get_rx_queue(channel);
661 efx_free_rx_buffers(rx_queue, rx_buf,
662 channel->rx_pkt_n_frags);
663 goto out;
664 }
665
666 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
667 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
668
669 if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
670 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
671 else
672 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
673out:
674 channel->rx_pkt_n_frags = 0;
675}
676
677int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
678{
679 struct efx_nic *efx = rx_queue->efx;
680 unsigned int entries;
681 int rc;
682
683 /* Create the smallest power-of-two aligned ring */
684 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
685 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
686 rx_queue->ptr_mask = entries - 1;
687
688 netif_dbg(efx, probe, efx->net_dev,
689 "creating RX queue %d size %#x mask %#x\n",
690 efx_rx_queue_index(rx_queue), efx->rxq_entries,
691 rx_queue->ptr_mask);
692
693 /* Allocate RX buffers */
694 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
695 GFP_KERNEL);
696 if (!rx_queue->buffer)
697 return -ENOMEM;
698
699 rc = efx_nic_probe_rx(rx_queue);
700 if (rc) {
701 kfree(rx_queue->buffer);
702 rx_queue->buffer = NULL;
703 }
704
705 return rc;
706}
707
708static void efx_init_rx_recycle_ring(struct efx_nic *efx,
709 struct efx_rx_queue *rx_queue)
710{
711 unsigned int bufs_in_recycle_ring, page_ring_size;
712
713 /* Set the RX recycle ring size */
714#ifdef CONFIG_PPC64
715 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
716#else
717 if (iommu_present(&pci_bus_type))
718 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
719 else
720 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
721#endif /* CONFIG_PPC64 */
722
723 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
724 efx->rx_bufs_per_page);
725 rx_queue->page_ring = kcalloc(page_ring_size,
726 sizeof(*rx_queue->page_ring), GFP_KERNEL);
727 rx_queue->page_ptr_mask = page_ring_size - 1;
728}
729
730void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
731{
732 struct efx_nic *efx = rx_queue->efx;
733 unsigned int max_fill, trigger, max_trigger;
734
735 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
736 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
737
738 /* Initialise ptr fields */
739 rx_queue->added_count = 0;
740 rx_queue->notified_count = 0;
741 rx_queue->removed_count = 0;
742 rx_queue->min_fill = -1U;
743 efx_init_rx_recycle_ring(efx, rx_queue);
744
745 rx_queue->page_remove = 0;
746 rx_queue->page_add = rx_queue->page_ptr_mask + 1;
747 rx_queue->page_recycle_count = 0;
748 rx_queue->page_recycle_failed = 0;
749 rx_queue->page_recycle_full = 0;
750
751 /* Initialise limit fields */
752 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
753 max_trigger =
754 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
755 if (rx_refill_threshold != 0) {
756 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
757 if (trigger > max_trigger)
758 trigger = max_trigger;
759 } else {
760 trigger = max_trigger;
761 }
762
763 rx_queue->max_fill = max_fill;
764 rx_queue->fast_fill_trigger = trigger;
765 rx_queue->refill_enabled = true;
766
767 /* Set up RX descriptor ring */
768 efx_nic_init_rx(rx_queue);
769}
770
771void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
772{
773 int i;
774 struct efx_nic *efx = rx_queue->efx;
775 struct efx_rx_buffer *rx_buf;
776
777 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
778 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
779
780 del_timer_sync(&rx_queue->slow_fill);
781
782 /* Release RX buffers from the current read ptr to the write ptr */
783 if (rx_queue->buffer) {
784 for (i = rx_queue->removed_count; i < rx_queue->added_count;
785 i++) {
786 unsigned index = i & rx_queue->ptr_mask;
787 rx_buf = efx_rx_buffer(rx_queue, index);
788 efx_fini_rx_buffer(rx_queue, rx_buf);
789 }
790 }
791
792 /* Unmap and release the pages in the recycle ring. Remove the ring. */
793 for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
794 struct page *page = rx_queue->page_ring[i];
795 struct efx_rx_page_state *state;
796
797 if (page == NULL)
798 continue;
799
800 state = page_address(page);
801 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
802 PAGE_SIZE << efx->rx_buffer_order,
803 DMA_FROM_DEVICE);
804 put_page(page);
805 }
806 kfree(rx_queue->page_ring);
807 rx_queue->page_ring = NULL;
808}
809
810void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
811{
812 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
813 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
814
815 efx_nic_remove_rx(rx_queue);
816
817 kfree(rx_queue->buffer);
818 rx_queue->buffer = NULL;
819}
820
821
822module_param(rx_refill_threshold, uint, 0444);
823MODULE_PARM_DESC(rx_refill_threshold,
824 "RX descriptor ring refill threshold (%)");
825
826#ifdef CONFIG_RFS_ACCEL
827
828static void efx_filter_rfs_work(struct work_struct *data)
829{
830 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
831 work);
832 struct efx_nic *efx = netdev_priv(req->net_dev);
833 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
834 int slot_idx = req - efx->rps_slot;
835 struct efx_arfs_rule *rule;
836 u16 arfs_id = 0;
837 int rc;
838
839 rc = efx->type->filter_insert(efx, &req->spec, true);
840 if (rc >= 0)
841 rc %= efx->type->max_rx_ip_filters;
842 if (efx->rps_hash_table) {
843 spin_lock_bh(&efx->rps_hash_lock);
844 rule = efx_rps_hash_find(efx, &req->spec);
845 /* The rule might have already gone, if someone else's request
846 * for the same spec was already worked and then expired before
847 * we got around to our work. In that case we have nothing
848 * tying us to an arfs_id, meaning that as soon as the filter
849 * is considered for expiry it will be removed.
850 */
851 if (rule) {
852 if (rc < 0)
853 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
854 else
855 rule->filter_id = rc;
856 arfs_id = rule->arfs_id;
857 }
858 spin_unlock_bh(&efx->rps_hash_lock);
859 }
860 if (rc >= 0) {
861 /* Remember this so we can check whether to expire the filter
862 * later.
863 */
864 mutex_lock(&efx->rps_mutex);
865 channel->rps_flow_id[rc] = req->flow_id;
866 ++channel->rfs_filters_added;
867 mutex_unlock(&efx->rps_mutex);
868
869 if (req->spec.ether_type == htons(ETH_P_IP))
870 netif_info(efx, rx_status, efx->net_dev,
871 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
872 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
873 req->spec.rem_host, ntohs(req->spec.rem_port),
874 req->spec.loc_host, ntohs(req->spec.loc_port),
875 req->rxq_index, req->flow_id, rc, arfs_id);
876 else
877 netif_info(efx, rx_status, efx->net_dev,
878 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
879 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
880 req->spec.rem_host, ntohs(req->spec.rem_port),
881 req->spec.loc_host, ntohs(req->spec.loc_port),
882 req->rxq_index, req->flow_id, rc, arfs_id);
883 }
884
885 /* Release references */
886 clear_bit(slot_idx, &efx->rps_slot_map);
887 dev_put(req->net_dev);
888}
889
890int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
891 u16 rxq_index, u32 flow_id)
892{
893 struct efx_nic *efx = netdev_priv(net_dev);
894 struct efx_async_filter_insertion *req;
895 struct efx_arfs_rule *rule;
896 struct flow_keys fk;
897 int slot_idx;
898 bool new;
899 int rc;
900
901 /* find a free slot */
902 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
903 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
904 break;
905 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
906 return -EBUSY;
907
908 if (flow_id == RPS_FLOW_ID_INVALID) {
909 rc = -EINVAL;
910 goto out_clear;
911 }
912
913 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
914 rc = -EPROTONOSUPPORT;
915 goto out_clear;
916 }
917
918 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
919 rc = -EPROTONOSUPPORT;
920 goto out_clear;
921 }
922 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
923 rc = -EPROTONOSUPPORT;
924 goto out_clear;
925 }
926
927 req = efx->rps_slot + slot_idx;
928 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
929 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
930 rxq_index);
931 req->spec.match_flags =
932 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
933 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
934 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
935 req->spec.ether_type = fk.basic.n_proto;
936 req->spec.ip_proto = fk.basic.ip_proto;
937
938 if (fk.basic.n_proto == htons(ETH_P_IP)) {
939 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
940 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
941 } else {
942 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
943 sizeof(struct in6_addr));
944 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
945 sizeof(struct in6_addr));
946 }
947
948 req->spec.rem_port = fk.ports.src;
949 req->spec.loc_port = fk.ports.dst;
950
951 if (efx->rps_hash_table) {
952 /* Add it to ARFS hash table */
953 spin_lock(&efx->rps_hash_lock);
954 rule = efx_rps_hash_add(efx, &req->spec, &new);
955 if (!rule) {
956 rc = -ENOMEM;
957 goto out_unlock;
958 }
959 if (new)
960 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
961 rc = rule->arfs_id;
962 /* Skip if existing or pending filter already does the right thing */
963 if (!new && rule->rxq_index == rxq_index &&
964 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
965 goto out_unlock;
966 rule->rxq_index = rxq_index;
967 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
968 spin_unlock(&efx->rps_hash_lock);
969 } else {
970 /* Without an ARFS hash table, we just use arfs_id 0 for all
971 * filters. This means if multiple flows hash to the same
972 * flow_id, all but the most recently touched will be eligible
973 * for expiry.
974 */
975 rc = 0;
976 }
977
978 /* Queue the request */
979 dev_hold(req->net_dev = net_dev);
980 INIT_WORK(&req->work, efx_filter_rfs_work);
981 req->rxq_index = rxq_index;
982 req->flow_id = flow_id;
983 schedule_work(&req->work);
984 return rc;
985out_unlock:
986 spin_unlock(&efx->rps_hash_lock);
987out_clear:
988 clear_bit(slot_idx, &efx->rps_slot_map);
989 return rc;
990}
991
992bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
993{
994 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
995 unsigned int channel_idx, index, size;
996 u32 flow_id;
997
998 if (!mutex_trylock(&efx->rps_mutex))
999 return false;
1000 expire_one = efx->type->filter_rfs_expire_one;
1001 channel_idx = efx->rps_expire_channel;
1002 index = efx->rps_expire_index;
1003 size = efx->type->max_rx_ip_filters;
1004 while (quota--) {
1005 struct efx_channel *channel = efx_get_channel(efx, channel_idx);
1006 flow_id = channel->rps_flow_id[index];
1007
1008 if (flow_id != RPS_FLOW_ID_INVALID &&
1009 expire_one(efx, flow_id, index)) {
1010 netif_info(efx, rx_status, efx->net_dev,
1011 "expired filter %d [queue %u flow %u]\n",
1012 index, channel_idx, flow_id);
1013 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1014 }
1015 if (++index == size) {
1016 if (++channel_idx == efx->n_channels)
1017 channel_idx = 0;
1018 index = 0;
1019 }
1020 }
1021 efx->rps_expire_channel = channel_idx;
1022 efx->rps_expire_index = index;
1023
1024 mutex_unlock(&efx->rps_mutex);
1025 return true;
1026}
1027
1028#endif /* CONFIG_RFS_ACCEL */
1029
1030/**
1031 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
1032 * @spec: Specification to test
1033 *
1034 * Return: %true if the specification is a non-drop RX filter that
1035 * matches a local MAC address I/G bit value of 1 or matches a local
1036 * IPv4 or IPv6 address value in the respective multicast address
1037 * range. Otherwise %false.
1038 */
1039bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
1040{
1041 if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
1042 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
1043 return false;
1044
1045 if (spec->match_flags &
1046 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
1047 is_multicast_ether_addr(spec->loc_mac))
1048 return true;
1049
1050 if ((spec->match_flags &
1051 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
1052 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
1053 if (spec->ether_type == htons(ETH_P_IP) &&
1054 ipv4_is_multicast(spec->loc_host[0]))
1055 return true;
1056 if (spec->ether_type == htons(ETH_P_IPV6) &&
1057 ((const u8 *)spec->loc_host)[0] == 0xff)
1058 return true;
1059 }
1060
1061 return false;
1062}