blob: 3c3d48e9ef4bbbf64a357d3eb3c8948b1dd03bbb [file] [log] [blame]
Mohsin Kazmie7cde312018-06-26 17:20:11 +02001/*
2 *------------------------------------------------------------------
3 * vhost-user-input
4 *
5 * Copyright (c) 2014-2018 Cisco and/or its affiliates.
6 * Licensed under the Apache License, Version 2.0 (the "License");
7 * you may not use this file except in compliance with the License.
8 * You may obtain a copy of the License at:
9 *
10 * http://www.apache.org/licenses/LICENSE-2.0
11 *
12 * Unless required by applicable law or agreed to in writing, software
13 * distributed under the License is distributed on an "AS IS" BASIS,
14 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 * See the License for the specific language governing permissions and
16 * limitations under the License.
17 *------------------------------------------------------------------
18 */
19
20#include <fcntl.h> /* for open */
21#include <sys/ioctl.h>
22#include <sys/socket.h>
23#include <sys/un.h>
24#include <sys/stat.h>
25#include <sys/types.h>
26#include <sys/uio.h> /* for iovec */
27#include <netinet/in.h>
28#include <sys/vfs.h>
29
30#include <linux/if_arp.h>
31#include <linux/if_tun.h>
32
33#include <vlib/vlib.h>
34#include <vlib/unix/unix.h>
35
36#include <vnet/ip/ip.h>
37
38#include <vnet/ethernet/ethernet.h>
39#include <vnet/devices/devices.h>
40#include <vnet/feature/feature.h>
41
42#include <vnet/devices/virtio/vhost_user.h>
43#include <vnet/devices/virtio/vhost_user_inline.h>
44
45/*
46 * When an RX queue is down but active, received packets
47 * must be discarded. This value controls up to how many
48 * packets will be discarded during each round.
49 */
50#define VHOST_USER_DOWN_DISCARD_COUNT 256
51
52/*
53 * When the number of available buffers gets under this threshold,
54 * RX node will start discarding packets.
55 */
56#define VHOST_USER_RX_BUFFER_STARVATION 32
57
58/*
59 * On the receive side, the host should free descriptors as soon
60 * as possible in order to avoid TX drop in the VM.
61 * This value controls the number of copy operations that are stacked
62 * before copy is done for all and descriptors are given back to
63 * the guest.
64 * The value 64 was obtained by testing (48 and 128 were not as good).
65 */
66#define VHOST_USER_RX_COPY_THRESHOLD 64
67
68vlib_node_registration_t vhost_user_input_node;
69
70#define foreach_vhost_user_input_func_error \
71 _(NO_ERROR, "no error") \
72 _(NO_BUFFER, "no available buffer") \
73 _(MMAP_FAIL, "mmap failure") \
74 _(INDIRECT_OVERFLOW, "indirect descriptor overflows table") \
75 _(UNDERSIZED_FRAME, "undersized ethernet frame received (< 14 bytes)") \
76 _(FULL_RX_QUEUE, "full rx queue (possible driver tx drop)")
77
78typedef enum
79{
80#define _(f,s) VHOST_USER_INPUT_FUNC_ERROR_##f,
81 foreach_vhost_user_input_func_error
82#undef _
83 VHOST_USER_INPUT_FUNC_N_ERROR,
84} vhost_user_input_func_error_t;
85
86static __clib_unused char *vhost_user_input_func_error_strings[] = {
87#define _(n,s) s,
88 foreach_vhost_user_input_func_error
89#undef _
90};
91
92static_always_inline void
93vhost_user_rx_trace (vhost_trace_t * t,
94 vhost_user_intf_t * vui, u16 qid,
Damjan Marionba1afaa2018-11-22 22:16:19 +010095 vlib_buffer_t * b, vhost_user_vring_t * txvq,
96 u16 last_avail_idx)
Mohsin Kazmie7cde312018-06-26 17:20:11 +020097{
98 vhost_user_main_t *vum = &vhost_user_main;
Mohsin Kazmie7cde312018-06-26 17:20:11 +020099 u32 desc_current = txvq->avail->ring[last_avail_idx & txvq->qsz_mask];
100 vring_desc_t *hdr_desc = 0;
101 virtio_net_hdr_mrg_rxbuf_t *hdr;
102 u32 hint = 0;
103
Dave Barachb7b92992018-10-17 10:38:51 -0400104 clib_memset (t, 0, sizeof (*t));
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200105 t->device_index = vui - vum->vhost_user_interfaces;
106 t->qid = qid;
107
108 hdr_desc = &txvq->desc[desc_current];
109 if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT)
110 {
111 t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT;
112 /* Header is the first here */
113 hdr_desc = map_guest_mem (vui, txvq->desc[desc_current].addr, &hint);
114 }
115 if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT)
116 {
117 t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED;
118 }
119 if (!(txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) &&
120 !(txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT))
121 {
122 t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC;
123 }
124
125 t->first_desc_len = hdr_desc ? hdr_desc->len : 0;
126
127 if (!hdr_desc || !(hdr = map_guest_mem (vui, hdr_desc->addr, &hint)))
128 {
129 t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_MAP_ERROR;
130 }
131 else
132 {
133 u32 len = vui->virtio_net_hdr_sz;
134 memcpy (&t->hdr, hdr, len > hdr_desc->len ? hdr_desc->len : len);
135 }
136}
137
138static_always_inline u32
139vhost_user_input_copy (vhost_user_intf_t * vui, vhost_copy_t * cpy,
140 u16 copy_len, u32 * map_hint)
141{
142 void *src0, *src1, *src2, *src3;
143 if (PREDICT_TRUE (copy_len >= 4))
144 {
145 if (PREDICT_FALSE (!(src2 = map_guest_mem (vui, cpy[0].src, map_hint))))
146 return 1;
147 if (PREDICT_FALSE (!(src3 = map_guest_mem (vui, cpy[1].src, map_hint))))
148 return 1;
149
150 while (PREDICT_TRUE (copy_len >= 4))
151 {
152 src0 = src2;
153 src1 = src3;
154
155 if (PREDICT_FALSE
156 (!(src2 = map_guest_mem (vui, cpy[2].src, map_hint))))
157 return 1;
158 if (PREDICT_FALSE
159 (!(src3 = map_guest_mem (vui, cpy[3].src, map_hint))))
160 return 1;
161
162 CLIB_PREFETCH (src2, 64, LOAD);
163 CLIB_PREFETCH (src3, 64, LOAD);
164
Dave Barach178cf492018-11-13 16:34:13 -0500165 clib_memcpy_fast ((void *) cpy[0].dst, src0, cpy[0].len);
166 clib_memcpy_fast ((void *) cpy[1].dst, src1, cpy[1].len);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200167 copy_len -= 2;
168 cpy += 2;
169 }
170 }
171 while (copy_len)
172 {
173 if (PREDICT_FALSE (!(src0 = map_guest_mem (vui, cpy->src, map_hint))))
174 return 1;
Dave Barach178cf492018-11-13 16:34:13 -0500175 clib_memcpy_fast ((void *) cpy->dst, src0, cpy->len);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200176 copy_len -= 1;
177 cpy += 1;
178 }
179 return 0;
180}
181
182/**
183 * Try to discard packets from the tx ring (VPP RX path).
184 * Returns the number of discarded packets.
185 */
186static_always_inline u32
187vhost_user_rx_discard_packet (vlib_main_t * vm,
188 vhost_user_intf_t * vui,
189 vhost_user_vring_t * txvq, u32 discard_max)
190{
191 /*
192 * On the RX side, each packet corresponds to one descriptor
193 * (it is the same whether it is a shallow descriptor, chained, or indirect).
194 * Therefore, discarding a packet is like discarding a descriptor.
195 */
196 u32 discarded_packets = 0;
197 u32 avail_idx = txvq->avail->idx;
Damjan Marionba1afaa2018-11-22 22:16:19 +0100198 u16 mask = txvq->qsz_mask;
199 u16 last_avail_idx = txvq->last_avail_idx;
200 u16 last_used_idx = txvq->last_used_idx;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200201 while (discarded_packets != discard_max)
202 {
203 if (avail_idx == txvq->last_avail_idx)
204 goto out;
205
Damjan Marionba1afaa2018-11-22 22:16:19 +0100206 u16 desc_chain_head = txvq->avail->ring[last_avail_idx & mask];
207 last_avail_idx++;
208 txvq->used->ring[last_used_idx & mask].id = desc_chain_head;
209 txvq->used->ring[last_used_idx & mask].len = 0;
210 vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & mask]);
211 last_used_idx++;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200212 discarded_packets++;
213 }
214
215out:
Damjan Marionba1afaa2018-11-22 22:16:19 +0100216 txvq->last_avail_idx = last_avail_idx;
217 txvq->last_used_idx = last_used_idx;
Damjan Marion96e8cd02018-11-23 14:56:55 +0100218 CLIB_MEMORY_STORE_BARRIER ();
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200219 txvq->used->idx = txvq->last_used_idx;
220 vhost_user_log_dirty_ring (vui, txvq, idx);
221 return discarded_packets;
222}
223
224/*
225 * In case of overflow, we need to rewind the array of allocated buffers.
226 */
Damjan Marion46bf8662018-11-22 22:25:38 +0100227static_always_inline void
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200228vhost_user_input_rewind_buffers (vlib_main_t * vm,
229 vhost_cpu_t * cpu, vlib_buffer_t * b_head)
230{
231 u32 bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
232 vlib_buffer_t *b_current = vlib_get_buffer (vm, bi_current);
233 b_current->current_length = 0;
234 b_current->flags = 0;
235 while (b_current != b_head)
236 {
237 cpu->rx_buffers_len++;
238 bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
239 b_current = vlib_get_buffer (vm, bi_current);
240 b_current->current_length = 0;
241 b_current->flags = 0;
242 }
243 cpu->rx_buffers_len++;
244}
245
Damjan Marion46bf8662018-11-22 22:25:38 +0100246static_always_inline u32
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200247vhost_user_if_input (vlib_main_t * vm,
248 vhost_user_main_t * vum,
249 vhost_user_intf_t * vui,
250 u16 qid, vlib_node_runtime_t * node,
251 vnet_hw_interface_rx_mode mode)
252{
253 vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)];
Damjan Marion9af45042018-11-21 09:51:42 +0100254 vnet_feature_main_t *fm = &feature_main;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200255 u16 n_rx_packets = 0;
256 u32 n_rx_bytes = 0;
257 u16 n_left;
258 u32 n_left_to_next, *to_next;
259 u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
260 u32 n_trace = vlib_get_trace_count (vm, node);
Damjan Marion8934a042019-02-09 23:29:26 +0100261 u32 buffer_data_size = vlib_buffer_get_default_data_size (vm);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200262 u32 map_hint = 0;
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100263 vhost_cpu_t *cpu = &vum->cpus[vm->thread_index];
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200264 u16 copy_len = 0;
Damjan Marion9af45042018-11-21 09:51:42 +0100265 u8 feature_arc_idx = fm->device_input_feature_arc_index;
266 u32 current_config_index = ~(u32) 0;
Damjan Marionba1afaa2018-11-22 22:16:19 +0100267 u16 mask = txvq->qsz_mask;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200268
Yichen Wang28812a02018-08-28 23:05:27 -0700269 /* The descriptor table is not ready yet */
270 if (PREDICT_FALSE (txvq->avail == 0))
Damjan Marionba1afaa2018-11-22 22:16:19 +0100271 goto done;
Yichen Wang28812a02018-08-28 23:05:27 -0700272
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200273 {
274 /* do we have pending interrupts ? */
275 vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)];
276 f64 now = vlib_time_now (vm);
277
278 if ((txvq->n_since_last_int) && (txvq->int_deadline < now))
279 vhost_user_send_call (vm, txvq);
280
281 if ((rxvq->n_since_last_int) && (rxvq->int_deadline < now))
282 vhost_user_send_call (vm, rxvq);
283 }
284
285 /*
286 * For adaptive mode, it is optimized to reduce interrupts.
287 * If the scheduler switches the input node to polling due
288 * to burst of traffic, we tell the driver no interrupt.
289 * When the traffic subsides, the scheduler switches the node back to
290 * interrupt mode. We must tell the driver we want interrupt.
291 */
292 if (PREDICT_FALSE (mode == VNET_HW_INTERFACE_RX_MODE_ADAPTIVE))
293 {
294 if ((node->flags &
295 VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) ||
296 !(node->flags &
297 VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE))
298 /* Tell driver we want notification */
299 txvq->used->flags = 0;
300 else
301 /* Tell driver we don't want notification */
302 txvq->used->flags = VRING_USED_F_NO_NOTIFY;
303 }
304
305 if (PREDICT_FALSE (txvq->avail->flags & 0xFFFE))
Damjan Marionba1afaa2018-11-22 22:16:19 +0100306 goto done;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200307
308 n_left = (u16) (txvq->avail->idx - txvq->last_avail_idx);
309
310 /* nothing to do */
311 if (PREDICT_FALSE (n_left == 0))
Damjan Marionba1afaa2018-11-22 22:16:19 +0100312 goto done;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200313
314 if (PREDICT_FALSE (!vui->admin_up || !(txvq->enabled)))
315 {
316 /*
317 * Discard input packet if interface is admin down or vring is not
318 * enabled.
319 * "For example, for a networking device, in the disabled state
320 * client must not supply any new RX packets, but must process
321 * and discard any TX packets."
322 */
323 vhost_user_rx_discard_packet (vm, vui, txvq,
324 VHOST_USER_DOWN_DISCARD_COUNT);
Damjan Marionba1afaa2018-11-22 22:16:19 +0100325 goto done;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200326 }
327
Damjan Marionba1afaa2018-11-22 22:16:19 +0100328 if (PREDICT_FALSE (n_left == (mask + 1)))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200329 {
330 /*
331 * Informational error logging when VPP is not
332 * receiving packets fast enough.
333 */
334 vlib_error_count (vm, node->node_index,
335 VHOST_USER_INPUT_FUNC_ERROR_FULL_RX_QUEUE, 1);
336 }
337
338 if (n_left > VLIB_FRAME_SIZE)
339 n_left = VLIB_FRAME_SIZE;
340
341 /*
342 * For small packets (<2kB), we will not need more than one vlib buffer
343 * per packet. In case packets are bigger, we will just yeld at some point
344 * in the loop and come back later. This is not an issue as for big packet,
345 * processing cost really comes from the memory copy.
346 * The assumption is that big packets will fit in 40 buffers.
347 */
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100348 if (PREDICT_FALSE (cpu->rx_buffers_len < n_left + 1 ||
349 cpu->rx_buffers_len < 40))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200350 {
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100351 u32 curr_len = cpu->rx_buffers_len;
352 cpu->rx_buffers_len +=
Damjan Marion671e60e2018-12-30 18:09:59 +0100353 vlib_buffer_alloc (vm, cpu->rx_buffers + curr_len,
354 VHOST_USER_RX_BUFFERS_N - curr_len);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200355
356 if (PREDICT_FALSE
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100357 (cpu->rx_buffers_len < VHOST_USER_RX_BUFFER_STARVATION))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200358 {
359 /* In case of buffer starvation, discard some packets from the queue
360 * and log the event.
361 * We keep doing best effort for the remaining packets. */
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100362 u32 flush = (n_left + 1 > cpu->rx_buffers_len) ?
363 n_left + 1 - cpu->rx_buffers_len : 1;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200364 flush = vhost_user_rx_discard_packet (vm, vui, txvq, flush);
365
366 n_left -= flush;
367 vlib_increment_simple_counter (vnet_main.
368 interface_main.sw_if_counters +
369 VNET_INTERFACE_COUNTER_DROP,
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100370 vm->thread_index, vui->sw_if_index,
371 flush);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200372
373 vlib_error_count (vm, vhost_user_input_node.index,
374 VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush);
375 }
376 }
377
Damjan Marion9af45042018-11-21 09:51:42 +0100378 if (PREDICT_FALSE (vnet_have_features (feature_arc_idx, vui->sw_if_index)))
379 {
380 vnet_feature_config_main_t *cm;
381 cm = &fm->feature_config_mains[feature_arc_idx];
382 current_config_index = vec_elt (cm->config_index_by_sw_if_index,
383 vui->sw_if_index);
384 vnet_get_config_data (&cm->config_main, &current_config_index,
385 &next_index, 0);
386 }
387
Damjan Marionba1afaa2018-11-22 22:16:19 +0100388 u16 last_avail_idx = txvq->last_avail_idx;
389 u16 last_used_idx = txvq->last_used_idx;
390
Damjan Marion92825382018-11-21 10:03:44 +0100391 vlib_get_new_next_frame (vm, node, next_index, to_next, n_left_to_next);
392
393 if (next_index == VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT)
394 {
395 /* give some hints to ethernet-input */
396 vlib_next_frame_t *nf;
397 vlib_frame_t *f;
398 ethernet_input_frame_t *ef;
399 nf = vlib_node_runtime_get_next_frame (vm, node, next_index);
400 f = vlib_get_frame (vm, nf->frame_index);
401 f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX;
402
403 ef = vlib_frame_scalar_args (f);
404 ef->sw_if_index = vui->sw_if_index;
405 ef->hw_if_index = vui->hw_if_index;
406 }
407
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200408 while (n_left > 0)
409 {
Damjan Marion92825382018-11-21 10:03:44 +0100410 vlib_buffer_t *b_head, *b_current;
411 u32 bi_current;
412 u16 desc_current;
413 u32 desc_data_offset;
414 vring_desc_t *desc_table = txvq->desc;
Damjan Marion6a8bfd42018-11-21 09:54:41 +0100415
Damjan Marion92825382018-11-21 10:03:44 +0100416 if (PREDICT_FALSE (cpu->rx_buffers_len <= 1))
Damjan Marion6a8bfd42018-11-21 09:54:41 +0100417 {
Damjan Marion92825382018-11-21 10:03:44 +0100418 /* Not enough rx_buffers
419 * Note: We yeld on 1 so we don't need to do an additional
420 * check for the next buffer prefetch.
421 */
422 n_left = 0;
423 break;
Damjan Marion6a8bfd42018-11-21 09:54:41 +0100424 }
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200425
Damjan Marionba1afaa2018-11-22 22:16:19 +0100426 desc_current = txvq->avail->ring[last_avail_idx & mask];
Damjan Marion92825382018-11-21 10:03:44 +0100427 cpu->rx_buffers_len--;
428 bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
429 b_head = b_current = vlib_get_buffer (vm, bi_current);
430 to_next[0] = bi_current; //We do that now so we can forget about bi_current
431 to_next++;
432 n_left_to_next--;
433
434 vlib_prefetch_buffer_with_index
435 (vm, cpu->rx_buffers[cpu->rx_buffers_len - 1], LOAD);
436
437 /* Just preset the used descriptor id and length for later */
Damjan Marionba1afaa2018-11-22 22:16:19 +0100438 txvq->used->ring[last_used_idx & mask].id = desc_current;
439 txvq->used->ring[last_used_idx & mask].len = 0;
440 vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & mask]);
Damjan Marion92825382018-11-21 10:03:44 +0100441
442 /* The buffer should already be initialized */
443 b_head->total_length_not_including_first_buffer = 0;
444 b_head->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
445
446 if (PREDICT_FALSE (n_trace))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200447 {
Damjan Marion92825382018-11-21 10:03:44 +0100448 //TODO: next_index is not exactly known at that point
449 vlib_trace_buffer (vm, node, next_index, b_head,
450 /* follow_chain */ 0);
451 vhost_trace_t *t0 =
452 vlib_add_trace (vm, node, b_head, sizeof (t0[0]));
Damjan Marionba1afaa2018-11-22 22:16:19 +0100453 vhost_user_rx_trace (t0, vui, qid, b_head, txvq, last_avail_idx);
Damjan Marion92825382018-11-21 10:03:44 +0100454 n_trace--;
455 vlib_set_trace_count (vm, node, n_trace);
456 }
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200457
Damjan Marion92825382018-11-21 10:03:44 +0100458 /* This depends on the setup but is very consistent
459 * So I think the CPU branch predictor will make a pretty good job
460 * at optimizing the decision. */
461 if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT)
462 {
463 desc_table = map_guest_mem (vui, txvq->desc[desc_current].addr,
464 &map_hint);
465 desc_current = 0;
466 if (PREDICT_FALSE (desc_table == 0))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200467 {
Damjan Marion92825382018-11-21 10:03:44 +0100468 vlib_error_count (vm, node->node_index,
469 VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
470 goto out;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200471 }
Damjan Marion92825382018-11-21 10:03:44 +0100472 }
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200473
Damjan Marion92825382018-11-21 10:03:44 +0100474 if (PREDICT_TRUE (vui->is_any_layout) ||
475 (!(desc_table[desc_current].flags & VIRTQ_DESC_F_NEXT)))
476 {
477 /* ANYLAYOUT or single buffer */
478 desc_data_offset = vui->virtio_net_hdr_sz;
479 }
480 else
481 {
482 /* CSR case without ANYLAYOUT, skip 1st buffer */
483 desc_data_offset = desc_table[desc_current].len;
484 }
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200485
Damjan Marion92825382018-11-21 10:03:44 +0100486 while (1)
487 {
488 /* Get more input if necessary. Or end of packet. */
489 if (desc_data_offset == desc_table[desc_current].len)
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200490 {
Damjan Marion92825382018-11-21 10:03:44 +0100491 if (PREDICT_FALSE (desc_table[desc_current].flags &
492 VIRTQ_DESC_F_NEXT))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200493 {
Damjan Marion92825382018-11-21 10:03:44 +0100494 desc_current = desc_table[desc_current].next;
495 desc_data_offset = 0;
496 }
497 else
498 {
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200499 goto out;
500 }
501 }
502
Damjan Marion92825382018-11-21 10:03:44 +0100503 /* Get more output if necessary. Or end of packet. */
Damjan Marion8934a042019-02-09 23:29:26 +0100504 if (PREDICT_FALSE (b_current->current_length == buffer_data_size))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200505 {
Damjan Marion92825382018-11-21 10:03:44 +0100506 if (PREDICT_FALSE (cpu->rx_buffers_len == 0))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200507 {
Damjan Marion92825382018-11-21 10:03:44 +0100508 /* Cancel speculation */
509 to_next--;
510 n_left_to_next++;
511
512 /*
513 * Checking if there are some left buffers.
514 * If not, just rewind the used buffers and stop.
515 * Note: Scheduled copies are not cancelled. This is
516 * not an issue as they would still be valid. Useless,
517 * but valid.
518 */
519 vhost_user_input_rewind_buffers (vm, cpu, b_head);
520 n_left = 0;
521 goto stop;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200522 }
523
Damjan Marion92825382018-11-21 10:03:44 +0100524 /* Get next output */
525 cpu->rx_buffers_len--;
526 u32 bi_next = cpu->rx_buffers[cpu->rx_buffers_len];
527 b_current->next_buffer = bi_next;
528 b_current->flags |= VLIB_BUFFER_NEXT_PRESENT;
529 bi_current = bi_next;
530 b_current = vlib_get_buffer (vm, bi_current);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200531 }
532
Damjan Marion92825382018-11-21 10:03:44 +0100533 /* Prepare a copy order executed later for the data */
534 vhost_copy_t *cpy = &cpu->copy[copy_len];
535 copy_len++;
536 u32 desc_data_l = desc_table[desc_current].len - desc_data_offset;
Damjan Marion8934a042019-02-09 23:29:26 +0100537 cpy->len = buffer_data_size - b_current->current_length;
Damjan Marion92825382018-11-21 10:03:44 +0100538 cpy->len = (cpy->len > desc_data_l) ? desc_data_l : cpy->len;
539 cpy->dst = (uword) (vlib_buffer_get_current (b_current) +
540 b_current->current_length);
541 cpy->src = desc_table[desc_current].addr + desc_data_offset;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200542
Damjan Marion92825382018-11-21 10:03:44 +0100543 desc_data_offset += cpy->len;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200544
Damjan Marion92825382018-11-21 10:03:44 +0100545 b_current->current_length += cpy->len;
546 b_head->total_length_not_including_first_buffer += cpy->len;
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200547 }
Damjan Marion92825382018-11-21 10:03:44 +0100548
549 out:
550
551 n_rx_bytes += b_head->total_length_not_including_first_buffer;
552 n_rx_packets++;
553
554 b_head->total_length_not_including_first_buffer -=
555 b_head->current_length;
556
557 /* consume the descriptor and return it as used */
Damjan Marionba1afaa2018-11-22 22:16:19 +0100558 last_avail_idx++;
559 last_used_idx++;
Damjan Marion92825382018-11-21 10:03:44 +0100560
561 VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b_head);
562
563 vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index;
564 vnet_buffer (b_head)->sw_if_index[VLIB_TX] = (u32) ~ 0;
565 b_head->error = 0;
566
567 if (current_config_index != ~(u32) 0)
568 {
569 b_head->current_config_index = current_config_index;
570 vnet_buffer (b_head)->feature_arc_index = feature_arc_idx;
571 }
572
573 n_left--;
574
575 /*
576 * Although separating memory copies from virtio ring parsing
577 * is beneficial, we can offer to perform the copies from time
578 * to time in order to free some space in the ring.
579 */
580 if (PREDICT_FALSE (copy_len >= VHOST_USER_RX_COPY_THRESHOLD))
581 {
582 if (PREDICT_FALSE (vhost_user_input_copy (vui, cpu->copy,
583 copy_len, &map_hint)))
584 {
585 vlib_error_count (vm, node->node_index,
586 VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
587 }
588 copy_len = 0;
589
590 /* give buffers back to driver */
Damjan Marion96e8cd02018-11-23 14:56:55 +0100591 CLIB_MEMORY_STORE_BARRIER ();
Damjan Marionba1afaa2018-11-22 22:16:19 +0100592 txvq->used->idx = last_used_idx;
Damjan Marion92825382018-11-21 10:03:44 +0100593 vhost_user_log_dirty_ring (vui, txvq, idx);
594 }
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200595 }
Damjan Marion92825382018-11-21 10:03:44 +0100596stop:
597 vlib_put_next_frame (vm, node, next_index, n_left_to_next);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200598
Damjan Marionba1afaa2018-11-22 22:16:19 +0100599 txvq->last_used_idx = last_used_idx;
600 txvq->last_avail_idx = last_avail_idx;
601
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200602 /* Do the memory copies */
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100603 if (PREDICT_FALSE (vhost_user_input_copy (vui, cpu->copy, copy_len,
604 &map_hint)))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200605 {
606 vlib_error_count (vm, node->node_index,
607 VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
608 }
609
610 /* give buffers back to driver */
Damjan Marion96e8cd02018-11-23 14:56:55 +0100611 CLIB_MEMORY_STORE_BARRIER ();
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200612 txvq->used->idx = txvq->last_used_idx;
613 vhost_user_log_dirty_ring (vui, txvq, idx);
614
615 /* interrupt (call) handling */
616 if ((txvq->callfd_idx != ~0) &&
617 !(txvq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
618 {
619 txvq->n_since_last_int += n_rx_packets;
620
621 if (txvq->n_since_last_int > vum->coalesce_frames)
622 vhost_user_send_call (vm, txvq);
623 }
624
625 /* increase rx counters */
626 vlib_increment_combined_counter
627 (vnet_main.interface_main.combined_sw_if_counters
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100628 + VNET_INTERFACE_COUNTER_RX, vm->thread_index, vui->sw_if_index,
629 n_rx_packets, n_rx_bytes);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200630
Damjan Marion7e0b17d2018-11-20 21:07:03 +0100631 vnet_device_increment_rx_packets (vm->thread_index, n_rx_packets);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200632
Damjan Marionba1afaa2018-11-22 22:16:19 +0100633done:
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200634 return n_rx_packets;
635}
636
637VLIB_NODE_FN (vhost_user_input_node) (vlib_main_t * vm,
638 vlib_node_runtime_t * node,
639 vlib_frame_t * frame)
640{
641 vhost_user_main_t *vum = &vhost_user_main;
642 uword n_rx_packets = 0;
643 vhost_user_intf_t *vui;
644 vnet_device_input_runtime_t *rt =
645 (vnet_device_input_runtime_t *) node->runtime_data;
646 vnet_device_and_queue_t *dq;
647
648 vec_foreach (dq, rt->devices_and_queues)
649 {
Sirshak Das5b718d52018-10-12 09:38:27 -0500650 if ((node->state == VLIB_NODE_STATE_POLLING) ||
651 clib_atomic_swap_acq_n (&dq->interrupt_pending, 0))
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200652 {
653 vui =
654 pool_elt_at_index (vum->vhost_user_interfaces, dq->dev_instance);
Damjan Marionbefe6912018-11-22 15:45:54 +0100655 n_rx_packets += vhost_user_if_input (vm, vum, vui, dq->queue_id, node,
656 dq->mode);
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200657 }
658 }
659
660 return n_rx_packets;
661}
662
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200663/* *INDENT-OFF* */
664VLIB_REGISTER_NODE (vhost_user_input_node) = {
665 .type = VLIB_NODE_TYPE_INPUT,
666 .name = "vhost-user-input",
667 .sibling_of = "device-input",
668
669 /* Will be enabled if/when hardware is detected. */
670 .state = VLIB_NODE_STATE_DISABLED,
671
672 .format_buffer = format_ethernet_header_with_length,
673 .format_trace = format_vhost_trace,
674
675 .n_errors = VHOST_USER_INPUT_FUNC_N_ERROR,
676 .error_strings = vhost_user_input_func_error_strings,
677};
678/* *INDENT-ON* */
Mohsin Kazmie7cde312018-06-26 17:20:11 +0200679
680/*
681 * fd.io coding-style-patch-verification: ON
682 *
683 * Local Variables:
684 * eval: (c-set-style "gnu")
685 * End:
686 */