Line data Source code
1 : /*
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/ethernet/ethernet.h>
37 : #include <vnet/devices/devices.h>
38 : #include <vnet/feature/feature.h>
39 : #include <vnet/udp/udp_packet.h>
40 : #include <vnet/tcp/tcp_packet.h>
41 : #include <vnet/interface/rx_queue_funcs.h>
42 :
43 : #include <vhost/vhost_user.h>
44 : #include <vhost/vhost_user_inline.h>
45 :
46 : #include <vnet/ip/ip4_packet.h>
47 : #include <vnet/ip/ip6_packet.h>
48 :
49 : /*
50 : * When an RX queue is down but active, received packets
51 : * must be discarded. This value controls up to how many
52 : * packets will be discarded during each round.
53 : */
54 : #define VHOST_USER_DOWN_DISCARD_COUNT 256
55 :
56 : /*
57 : * When the number of available buffers gets under this threshold,
58 : * RX node will start discarding packets.
59 : */
60 : #define VHOST_USER_RX_BUFFER_STARVATION 32
61 :
62 : /*
63 : * On the receive side, the host should free descriptors as soon
64 : * as possible in order to avoid TX drop in the VM.
65 : * This value controls the number of copy operations that are stacked
66 : * before copy is done for all and descriptors are given back to
67 : * the guest.
68 : * The value 64 was obtained by testing (48 and 128 were not as good).
69 : */
70 : #define VHOST_USER_RX_COPY_THRESHOLD 64
71 :
72 : extern vlib_node_registration_t vhost_user_input_node;
73 :
74 : #define foreach_vhost_user_input_func_error \
75 : _(NO_ERROR, "no error") \
76 : _(NO_BUFFER, "no available buffer") \
77 : _(MMAP_FAIL, "mmap failure") \
78 : _(INDIRECT_OVERFLOW, "indirect descriptor overflows table") \
79 : _(UNDERSIZED_FRAME, "undersized ethernet frame received (< 14 bytes)") \
80 : _(NOT_READY, "vhost interface not ready or down") \
81 : _(FULL_RX_QUEUE, "full rx queue (possible driver tx drop)")
82 :
83 : typedef enum
84 : {
85 : #define _(f,s) VHOST_USER_INPUT_FUNC_ERROR_##f,
86 : foreach_vhost_user_input_func_error
87 : #undef _
88 : VHOST_USER_INPUT_FUNC_N_ERROR,
89 : } vhost_user_input_func_error_t;
90 :
91 : static __clib_unused char *vhost_user_input_func_error_strings[] = {
92 : #define _(n,s) s,
93 : foreach_vhost_user_input_func_error
94 : #undef _
95 : };
96 :
97 : static_always_inline void
98 0 : vhost_user_rx_trace (vhost_trace_t * t,
99 : vhost_user_intf_t * vui, u16 qid,
100 : vlib_buffer_t * b, vhost_user_vring_t * txvq,
101 : u16 last_avail_idx)
102 : {
103 0 : vhost_user_main_t *vum = &vhost_user_main;
104 0 : u32 desc_current = txvq->avail->ring[last_avail_idx & txvq->qsz_mask];
105 0 : vnet_virtio_vring_desc_t *hdr_desc = 0;
106 : vnet_virtio_net_hdr_mrg_rxbuf_t *hdr;
107 0 : u32 hint = 0;
108 :
109 0 : clib_memset (t, 0, sizeof (*t));
110 0 : t->device_index = vui - vum->vhost_user_interfaces;
111 0 : t->qid = qid;
112 :
113 0 : hdr_desc = &txvq->desc[desc_current];
114 0 : if (txvq->desc[desc_current].flags & VRING_DESC_F_INDIRECT)
115 : {
116 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT;
117 : /* Header is the first here */
118 0 : hdr_desc = map_guest_mem (vui, txvq->desc[desc_current].addr, &hint);
119 : }
120 0 : if (txvq->desc[desc_current].flags & VRING_DESC_F_NEXT)
121 : {
122 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED;
123 : }
124 0 : if (!(txvq->desc[desc_current].flags & VRING_DESC_F_NEXT) &&
125 0 : !(txvq->desc[desc_current].flags & VRING_DESC_F_INDIRECT))
126 : {
127 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC;
128 : }
129 :
130 0 : t->first_desc_len = hdr_desc ? hdr_desc->len : 0;
131 :
132 0 : if (!hdr_desc || !(hdr = map_guest_mem (vui, hdr_desc->addr, &hint)))
133 : {
134 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_MAP_ERROR;
135 : }
136 : else
137 : {
138 0 : u32 len = vui->virtio_net_hdr_sz;
139 0 : memcpy (&t->hdr, hdr, len > hdr_desc->len ? hdr_desc->len : len);
140 : }
141 0 : }
142 :
143 : static_always_inline u32
144 0 : vhost_user_input_copy (vhost_user_intf_t * vui, vhost_copy_t * cpy,
145 : u16 copy_len, u32 * map_hint)
146 : {
147 : void *src0, *src1, *src2, *src3;
148 0 : if (PREDICT_TRUE (copy_len >= 4))
149 : {
150 0 : if (PREDICT_FALSE (!(src2 = map_guest_mem (vui, cpy[0].src, map_hint))))
151 0 : return 1;
152 0 : if (PREDICT_FALSE (!(src3 = map_guest_mem (vui, cpy[1].src, map_hint))))
153 0 : return 1;
154 :
155 0 : while (PREDICT_TRUE (copy_len >= 4))
156 : {
157 0 : src0 = src2;
158 0 : src1 = src3;
159 :
160 0 : if (PREDICT_FALSE
161 : (!(src2 = map_guest_mem (vui, cpy[2].src, map_hint))))
162 0 : return 1;
163 0 : if (PREDICT_FALSE
164 : (!(src3 = map_guest_mem (vui, cpy[3].src, map_hint))))
165 0 : return 1;
166 :
167 0 : clib_prefetch_load (src2);
168 0 : clib_prefetch_load (src3);
169 :
170 0 : clib_memcpy_fast ((void *) cpy[0].dst, src0, cpy[0].len);
171 0 : clib_memcpy_fast ((void *) cpy[1].dst, src1, cpy[1].len);
172 0 : copy_len -= 2;
173 0 : cpy += 2;
174 : }
175 : }
176 0 : while (copy_len)
177 : {
178 0 : if (PREDICT_FALSE (!(src0 = map_guest_mem (vui, cpy->src, map_hint))))
179 0 : return 1;
180 0 : clib_memcpy_fast ((void *) cpy->dst, src0, cpy->len);
181 0 : copy_len -= 1;
182 0 : cpy += 1;
183 : }
184 0 : return 0;
185 : }
186 :
187 : /**
188 : * Try to discard packets from the tx ring (VPP RX path).
189 : * Returns the number of discarded packets.
190 : */
191 : static_always_inline u32
192 0 : vhost_user_rx_discard_packet (vlib_main_t * vm,
193 : vhost_user_intf_t * vui,
194 : vhost_user_vring_t * txvq, u32 discard_max)
195 : {
196 : /*
197 : * On the RX side, each packet corresponds to one descriptor
198 : * (it is the same whether it is a shallow descriptor, chained, or indirect).
199 : * Therefore, discarding a packet is like discarding a descriptor.
200 : */
201 0 : u32 discarded_packets = 0;
202 0 : u32 avail_idx = txvq->avail->idx;
203 0 : u16 mask = txvq->qsz_mask;
204 0 : u16 last_avail_idx = txvq->last_avail_idx;
205 0 : u16 last_used_idx = txvq->last_used_idx;
206 0 : while (discarded_packets != discard_max)
207 : {
208 0 : if (avail_idx == last_avail_idx)
209 0 : goto out;
210 :
211 0 : u16 desc_chain_head = txvq->avail->ring[last_avail_idx & mask];
212 0 : last_avail_idx++;
213 0 : txvq->used->ring[last_used_idx & mask].id = desc_chain_head;
214 0 : txvq->used->ring[last_used_idx & mask].len = 0;
215 0 : vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & mask]);
216 0 : last_used_idx++;
217 0 : discarded_packets++;
218 : }
219 :
220 0 : out:
221 0 : txvq->last_avail_idx = last_avail_idx;
222 0 : txvq->last_used_idx = last_used_idx;
223 0 : CLIB_MEMORY_STORE_BARRIER ();
224 0 : txvq->used->idx = txvq->last_used_idx;
225 0 : vhost_user_log_dirty_ring (vui, txvq, idx);
226 0 : return discarded_packets;
227 : }
228 :
229 : /*
230 : * In case of overflow, we need to rewind the array of allocated buffers.
231 : */
232 : static_always_inline void
233 0 : vhost_user_input_rewind_buffers (vlib_main_t * vm,
234 : vhost_cpu_t * cpu, vlib_buffer_t * b_head)
235 : {
236 0 : u32 bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
237 0 : vlib_buffer_t *b_current = vlib_get_buffer (vm, bi_current);
238 0 : b_current->current_length = 0;
239 0 : b_current->flags = 0;
240 0 : while (b_current != b_head)
241 : {
242 0 : cpu->rx_buffers_len++;
243 0 : bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
244 0 : b_current = vlib_get_buffer (vm, bi_current);
245 0 : b_current->current_length = 0;
246 0 : b_current->flags = 0;
247 : }
248 0 : cpu->rx_buffers_len++;
249 0 : }
250 :
251 : static_always_inline void
252 0 : vhost_user_handle_rx_offload (vlib_buffer_t *b0, u8 *b0_data,
253 : vnet_virtio_net_hdr_t *hdr)
254 : {
255 0 : u8 l4_hdr_sz = 0;
256 0 : u8 l4_proto = 0;
257 0 : ethernet_header_t *eh = (ethernet_header_t *) b0_data;
258 0 : u16 ethertype = clib_net_to_host_u16 (eh->type);
259 0 : u16 l2hdr_sz = sizeof (ethernet_header_t);
260 0 : vnet_buffer_oflags_t oflags = 0;
261 :
262 0 : if (ethernet_frame_is_tagged (ethertype))
263 : {
264 0 : ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (eh + 1);
265 :
266 0 : ethertype = clib_net_to_host_u16 (vlan->type);
267 0 : l2hdr_sz += sizeof (*vlan);
268 0 : if (ethertype == ETHERNET_TYPE_VLAN)
269 : {
270 0 : vlan++;
271 0 : ethertype = clib_net_to_host_u16 (vlan->type);
272 0 : l2hdr_sz += sizeof (*vlan);
273 : }
274 : }
275 0 : vnet_buffer (b0)->l2_hdr_offset = 0;
276 0 : vnet_buffer (b0)->l3_hdr_offset = l2hdr_sz;
277 0 : vnet_buffer (b0)->l4_hdr_offset = hdr->csum_start;
278 0 : b0->flags |= (VNET_BUFFER_F_L2_HDR_OFFSET_VALID |
279 : VNET_BUFFER_F_L3_HDR_OFFSET_VALID |
280 : VNET_BUFFER_F_L4_HDR_OFFSET_VALID);
281 :
282 0 : if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP4))
283 : {
284 0 : ip4_header_t *ip4 = (ip4_header_t *) (b0_data + l2hdr_sz);
285 0 : l4_proto = ip4->protocol;
286 0 : b0->flags |= VNET_BUFFER_F_IS_IP4;
287 0 : oflags |= VNET_BUFFER_OFFLOAD_F_IP_CKSUM;
288 : }
289 0 : else if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP6))
290 : {
291 0 : ip6_header_t *ip6 = (ip6_header_t *) (b0_data + l2hdr_sz);
292 0 : l4_proto = ip6->protocol;
293 0 : b0->flags |= VNET_BUFFER_F_IS_IP6;
294 : }
295 :
296 0 : if (l4_proto == IP_PROTOCOL_TCP)
297 : {
298 0 : tcp_header_t *tcp = (tcp_header_t *)
299 0 : (b0_data + vnet_buffer (b0)->l4_hdr_offset);
300 0 : l4_hdr_sz = tcp_header_bytes (tcp);
301 0 : oflags |= VNET_BUFFER_OFFLOAD_F_TCP_CKSUM;
302 : }
303 0 : else if (l4_proto == IP_PROTOCOL_UDP)
304 : {
305 0 : l4_hdr_sz = sizeof (udp_header_t);
306 0 : oflags |= VNET_BUFFER_OFFLOAD_F_UDP_CKSUM;
307 : }
308 :
309 0 : if (hdr->gso_type == VIRTIO_NET_HDR_GSO_UDP)
310 : {
311 0 : vnet_buffer2 (b0)->gso_size = hdr->gso_size;
312 0 : vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz;
313 0 : b0->flags |= VNET_BUFFER_F_GSO;
314 : }
315 0 : else if (hdr->gso_type == VIRTIO_NET_HDR_GSO_TCPV4)
316 : {
317 0 : vnet_buffer2 (b0)->gso_size = hdr->gso_size;
318 0 : vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz;
319 0 : b0->flags |= (VNET_BUFFER_F_GSO | VNET_BUFFER_F_IS_IP4);
320 : }
321 0 : else if (hdr->gso_type == VIRTIO_NET_HDR_GSO_TCPV6)
322 : {
323 0 : vnet_buffer2 (b0)->gso_size = hdr->gso_size;
324 0 : vnet_buffer2 (b0)->gso_l4_hdr_sz = l4_hdr_sz;
325 0 : b0->flags |= (VNET_BUFFER_F_GSO | VNET_BUFFER_F_IS_IP6);
326 : }
327 :
328 0 : if (oflags)
329 0 : vnet_buffer_offload_flags_set (b0, oflags);
330 0 : }
331 :
332 : static_always_inline void
333 0 : vhost_user_input_do_interrupt (vlib_main_t * vm, vhost_user_intf_t * vui,
334 : vhost_user_vring_t * txvq,
335 : vhost_user_vring_t * rxvq)
336 : {
337 0 : f64 now = vlib_time_now (vm);
338 :
339 0 : if ((txvq->n_since_last_int) && (txvq->int_deadline < now))
340 0 : vhost_user_send_call (vm, vui, txvq);
341 :
342 0 : if ((rxvq->n_since_last_int) && (rxvq->int_deadline < now))
343 0 : vhost_user_send_call (vm, vui, rxvq);
344 0 : }
345 :
346 : static_always_inline void
347 0 : vhost_user_input_setup_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
348 : vhost_user_intf_t * vui,
349 : u32 * current_config_index, u32 * next_index,
350 : u32 ** to_next, u32 * n_left_to_next)
351 : {
352 0 : vnet_feature_main_t *fm = &feature_main;
353 0 : u8 feature_arc_idx = fm->device_input_feature_arc_index;
354 :
355 0 : if (PREDICT_FALSE (vnet_have_features (feature_arc_idx, vui->sw_if_index)))
356 : {
357 : vnet_feature_config_main_t *cm;
358 0 : cm = &fm->feature_config_mains[feature_arc_idx];
359 0 : *current_config_index = vec_elt (cm->config_index_by_sw_if_index,
360 : vui->sw_if_index);
361 0 : vnet_get_config_data (&cm->config_main, current_config_index,
362 : next_index, 0);
363 : }
364 :
365 0 : vlib_get_new_next_frame (vm, node, *next_index, *to_next, *n_left_to_next);
366 :
367 0 : if (*next_index == VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT)
368 : {
369 : /* give some hints to ethernet-input */
370 : vlib_next_frame_t *nf;
371 : vlib_frame_t *f;
372 : ethernet_input_frame_t *ef;
373 0 : nf = vlib_node_runtime_get_next_frame (vm, node, *next_index);
374 0 : f = vlib_get_frame (vm, nf->frame);
375 0 : f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX;
376 :
377 0 : ef = vlib_frame_scalar_args (f);
378 0 : ef->sw_if_index = vui->sw_if_index;
379 0 : ef->hw_if_index = vui->hw_if_index;
380 0 : vlib_frame_no_append (f);
381 : }
382 0 : }
383 :
384 : static_always_inline u32
385 161953 : vhost_user_if_input (vlib_main_t *vm, vhost_user_main_t *vum,
386 : vhost_user_intf_t *vui, u16 qid,
387 : vlib_node_runtime_t *node, u8 enable_csum)
388 : {
389 161953 : vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)];
390 161953 : vnet_feature_main_t *fm = &feature_main;
391 161953 : u16 n_rx_packets = 0;
392 161953 : u32 n_rx_bytes = 0;
393 : u16 n_left;
394 : u32 n_left_to_next, *to_next;
395 161953 : u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
396 161953 : u32 n_trace = vlib_get_trace_count (vm, node);
397 161953 : u32 buffer_data_size = vlib_buffer_get_default_data_size (vm);
398 161953 : u32 map_hint = 0;
399 161953 : vhost_cpu_t *cpu = &vum->cpus[vm->thread_index];
400 161953 : u16 copy_len = 0;
401 161953 : u8 feature_arc_idx = fm->device_input_feature_arc_index;
402 161953 : u32 current_config_index = ~(u32) 0;
403 161953 : u16 mask = txvq->qsz_mask;
404 :
405 : /* The descriptor table is not ready yet */
406 161953 : if (PREDICT_FALSE (txvq->avail == 0))
407 161953 : goto done;
408 :
409 : {
410 : /* do we have pending interrupts ? */
411 0 : vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)];
412 0 : vhost_user_input_do_interrupt (vm, vui, txvq, rxvq);
413 : }
414 :
415 : /*
416 : * For adaptive mode, it is optimized to reduce interrupts.
417 : * If the scheduler switches the input node to polling due
418 : * to burst of traffic, we tell the driver no interrupt.
419 : * When the traffic subsides, the scheduler switches the node back to
420 : * interrupt mode. We must tell the driver we want interrupt.
421 : */
422 0 : if (PREDICT_FALSE (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE))
423 : {
424 0 : if ((node->flags &
425 0 : VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) ||
426 0 : !(node->flags &
427 : VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE))
428 : /* Tell driver we want notification */
429 0 : txvq->used->flags = 0;
430 : else
431 : /* Tell driver we don't want notification */
432 0 : txvq->used->flags = VRING_USED_F_NO_NOTIFY;
433 : }
434 :
435 0 : if (PREDICT_FALSE (txvq->avail->flags & 0xFFFE))
436 0 : goto done;
437 :
438 0 : n_left = (u16) (txvq->avail->idx - txvq->last_avail_idx);
439 :
440 : /* nothing to do */
441 0 : if (PREDICT_FALSE (n_left == 0))
442 0 : goto done;
443 :
444 0 : if (PREDICT_FALSE (!vui->admin_up || !(txvq->enabled)))
445 : {
446 : /*
447 : * Discard input packet if interface is admin down or vring is not
448 : * enabled.
449 : * "For example, for a networking device, in the disabled state
450 : * client must not supply any new RX packets, but must process
451 : * and discard any TX packets."
452 : */
453 0 : vhost_user_rx_discard_packet (vm, vui, txvq,
454 : VHOST_USER_DOWN_DISCARD_COUNT);
455 0 : goto done;
456 : }
457 :
458 0 : if (PREDICT_FALSE (n_left == (mask + 1)))
459 : {
460 : /*
461 : * Informational error logging when VPP is not
462 : * receiving packets fast enough.
463 : */
464 0 : vlib_error_count (vm, node->node_index,
465 : VHOST_USER_INPUT_FUNC_ERROR_FULL_RX_QUEUE, 1);
466 : }
467 :
468 0 : if (n_left > VLIB_FRAME_SIZE)
469 0 : n_left = VLIB_FRAME_SIZE;
470 :
471 : /*
472 : * For small packets (<2kB), we will not need more than one vlib buffer
473 : * per packet. In case packets are bigger, we will just yield at some point
474 : * in the loop and come back later. This is not an issue as for big packet,
475 : * processing cost really comes from the memory copy.
476 : * The assumption is that big packets will fit in 40 buffers.
477 : */
478 0 : if (PREDICT_FALSE (cpu->rx_buffers_len < n_left + 1 ||
479 : cpu->rx_buffers_len < 40))
480 : {
481 0 : u32 curr_len = cpu->rx_buffers_len;
482 0 : cpu->rx_buffers_len +=
483 0 : vlib_buffer_alloc (vm, cpu->rx_buffers + curr_len,
484 : VHOST_USER_RX_BUFFERS_N - curr_len);
485 :
486 0 : if (PREDICT_FALSE
487 : (cpu->rx_buffers_len < VHOST_USER_RX_BUFFER_STARVATION))
488 : {
489 : /* In case of buffer starvation, discard some packets from the queue
490 : * and log the event.
491 : * We keep doing best effort for the remaining packets. */
492 0 : u32 flush = (n_left + 1 > cpu->rx_buffers_len) ?
493 0 : n_left + 1 - cpu->rx_buffers_len : 1;
494 0 : flush = vhost_user_rx_discard_packet (vm, vui, txvq, flush);
495 :
496 0 : n_left -= flush;
497 0 : vlib_increment_simple_counter (vnet_main.
498 : interface_main.sw_if_counters +
499 : VNET_INTERFACE_COUNTER_DROP,
500 : vm->thread_index, vui->sw_if_index,
501 : flush);
502 :
503 0 : vlib_error_count (vm, vhost_user_input_node.index,
504 : VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush);
505 : }
506 : }
507 :
508 0 : vhost_user_input_setup_frame (vm, node, vui, ¤t_config_index,
509 : &next_index, &to_next, &n_left_to_next);
510 :
511 0 : u16 last_avail_idx = txvq->last_avail_idx;
512 0 : u16 last_used_idx = txvq->last_used_idx;
513 :
514 0 : while (n_left > 0)
515 : {
516 : vlib_buffer_t *b_head, *b_current;
517 : u32 bi_current;
518 : u16 desc_current;
519 : u32 desc_data_offset;
520 0 : vnet_virtio_vring_desc_t *desc_table = txvq->desc;
521 :
522 0 : if (PREDICT_FALSE (cpu->rx_buffers_len <= 1))
523 : {
524 : /* Not enough rx_buffers
525 : * Note: We yeld on 1 so we don't need to do an additional
526 : * check for the next buffer prefetch.
527 : */
528 0 : n_left = 0;
529 0 : break;
530 : }
531 :
532 0 : desc_current = txvq->avail->ring[last_avail_idx & mask];
533 0 : cpu->rx_buffers_len--;
534 0 : bi_current = cpu->rx_buffers[cpu->rx_buffers_len];
535 0 : b_head = b_current = vlib_get_buffer (vm, bi_current);
536 0 : to_next[0] = bi_current; //We do that now so we can forget about bi_current
537 0 : to_next++;
538 0 : n_left_to_next--;
539 :
540 0 : vlib_prefetch_buffer_with_index
541 : (vm, cpu->rx_buffers[cpu->rx_buffers_len - 1], LOAD);
542 :
543 : /* Just preset the used descriptor id and length for later */
544 0 : txvq->used->ring[last_used_idx & mask].id = desc_current;
545 0 : txvq->used->ring[last_used_idx & mask].len = 0;
546 0 : vhost_user_log_dirty_ring (vui, txvq, ring[last_used_idx & mask]);
547 :
548 : /* The buffer should already be initialized */
549 0 : b_head->total_length_not_including_first_buffer = 0;
550 0 : b_head->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
551 :
552 0 : if (PREDICT_FALSE
553 : (n_trace > 0 && vlib_trace_buffer (vm, node, next_index, b_head,
554 : /* follow_chain */ 0)))
555 : {
556 : vhost_trace_t *t0 =
557 0 : vlib_add_trace (vm, node, b_head, sizeof (t0[0]));
558 0 : vhost_user_rx_trace (t0, vui, qid, b_head, txvq, last_avail_idx);
559 0 : n_trace--;
560 0 : vlib_set_trace_count (vm, node, n_trace);
561 : }
562 :
563 : /* This depends on the setup but is very consistent
564 : * So I think the CPU branch predictor will make a pretty good job
565 : * at optimizing the decision. */
566 0 : if (txvq->desc[desc_current].flags & VRING_DESC_F_INDIRECT)
567 : {
568 0 : desc_table = map_guest_mem (vui, txvq->desc[desc_current].addr,
569 : &map_hint);
570 0 : desc_current = 0;
571 0 : if (PREDICT_FALSE (desc_table == 0))
572 : {
573 0 : vlib_error_count (vm, node->node_index,
574 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
575 0 : goto out;
576 : }
577 : }
578 :
579 0 : desc_data_offset = vui->virtio_net_hdr_sz;
580 :
581 0 : if (enable_csum)
582 : {
583 : vnet_virtio_net_hdr_mrg_rxbuf_t *hdr;
584 : u8 *b_data;
585 : u16 current;
586 :
587 0 : hdr = map_guest_mem (vui, desc_table[desc_current].addr, &map_hint);
588 0 : if (PREDICT_FALSE (hdr == 0))
589 : {
590 0 : vlib_error_count (vm, node->node_index,
591 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
592 0 : goto out;
593 : }
594 0 : if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
595 : {
596 0 : if ((desc_data_offset == desc_table[desc_current].len) &&
597 0 : (desc_table[desc_current].flags & VRING_DESC_F_NEXT))
598 : {
599 0 : current = desc_table[desc_current].next;
600 0 : b_data = map_guest_mem (vui, desc_table[current].addr,
601 : &map_hint);
602 0 : if (PREDICT_FALSE (b_data == 0))
603 : {
604 0 : vlib_error_count (vm, node->node_index,
605 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL,
606 : 1);
607 0 : goto out;
608 : }
609 : }
610 : else
611 0 : b_data = (u8 *) hdr + desc_data_offset;
612 :
613 0 : vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr);
614 : }
615 : }
616 :
617 : while (1)
618 0 : {
619 : /* Get more input if necessary. Or end of packet. */
620 0 : if (desc_data_offset == desc_table[desc_current].len)
621 : {
622 0 : if (PREDICT_FALSE (desc_table[desc_current].flags &
623 : VRING_DESC_F_NEXT))
624 : {
625 0 : desc_current = desc_table[desc_current].next;
626 0 : desc_data_offset = 0;
627 : }
628 : else
629 : {
630 0 : goto out;
631 : }
632 : }
633 :
634 : /* Get more output if necessary. Or end of packet. */
635 0 : if (PREDICT_FALSE (b_current->current_length == buffer_data_size))
636 : {
637 0 : if (PREDICT_FALSE (cpu->rx_buffers_len == 0))
638 : {
639 : /* Cancel speculation */
640 0 : to_next--;
641 0 : n_left_to_next++;
642 :
643 : /*
644 : * Checking if there are some left buffers.
645 : * If not, just rewind the used buffers and stop.
646 : * Note: Scheduled copies are not cancelled. This is
647 : * not an issue as they would still be valid. Useless,
648 : * but valid.
649 : */
650 0 : vhost_user_input_rewind_buffers (vm, cpu, b_head);
651 0 : n_left = 0;
652 0 : goto stop;
653 : }
654 :
655 : /* Get next output */
656 0 : cpu->rx_buffers_len--;
657 0 : u32 bi_next = cpu->rx_buffers[cpu->rx_buffers_len];
658 0 : b_current->next_buffer = bi_next;
659 0 : b_current->flags |= VLIB_BUFFER_NEXT_PRESENT;
660 0 : bi_current = bi_next;
661 0 : b_current = vlib_get_buffer (vm, bi_current);
662 : }
663 :
664 : /* Prepare a copy order executed later for the data */
665 0 : ASSERT (copy_len < VHOST_USER_COPY_ARRAY_N);
666 0 : vhost_copy_t *cpy = &cpu->copy[copy_len];
667 0 : copy_len++;
668 0 : u32 desc_data_l = desc_table[desc_current].len - desc_data_offset;
669 0 : cpy->len = buffer_data_size - b_current->current_length;
670 0 : cpy->len = (cpy->len > desc_data_l) ? desc_data_l : cpy->len;
671 0 : cpy->dst = (uword) (vlib_buffer_get_current (b_current) +
672 0 : b_current->current_length);
673 0 : cpy->src = desc_table[desc_current].addr + desc_data_offset;
674 :
675 0 : desc_data_offset += cpy->len;
676 :
677 0 : b_current->current_length += cpy->len;
678 0 : b_head->total_length_not_including_first_buffer += cpy->len;
679 : }
680 :
681 0 : out:
682 :
683 0 : n_rx_bytes += b_head->total_length_not_including_first_buffer;
684 0 : n_rx_packets++;
685 :
686 0 : b_head->total_length_not_including_first_buffer -=
687 0 : b_head->current_length;
688 :
689 : /* consume the descriptor and return it as used */
690 0 : last_avail_idx++;
691 0 : last_used_idx++;
692 :
693 0 : vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index;
694 0 : vnet_buffer (b_head)->sw_if_index[VLIB_TX] = (u32) ~ 0;
695 0 : b_head->error = 0;
696 :
697 0 : if (current_config_index != ~(u32) 0)
698 : {
699 0 : b_head->current_config_index = current_config_index;
700 0 : vnet_buffer (b_head)->feature_arc_index = feature_arc_idx;
701 : }
702 :
703 0 : n_left--;
704 :
705 : /*
706 : * Although separating memory copies from virtio ring parsing
707 : * is beneficial, we can offer to perform the copies from time
708 : * to time in order to free some space in the ring.
709 : */
710 0 : if (PREDICT_FALSE (copy_len >= VHOST_USER_RX_COPY_THRESHOLD))
711 : {
712 0 : if (PREDICT_FALSE (vhost_user_input_copy (vui, cpu->copy,
713 : copy_len, &map_hint)))
714 : {
715 0 : vlib_error_count (vm, node->node_index,
716 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
717 : }
718 0 : copy_len = 0;
719 :
720 : /* give buffers back to driver */
721 0 : CLIB_MEMORY_STORE_BARRIER ();
722 0 : txvq->used->idx = last_used_idx;
723 0 : vhost_user_log_dirty_ring (vui, txvq, idx);
724 : }
725 : }
726 0 : stop:
727 0 : vlib_put_next_frame (vm, node, next_index, n_left_to_next);
728 :
729 0 : txvq->last_used_idx = last_used_idx;
730 0 : txvq->last_avail_idx = last_avail_idx;
731 :
732 : /* Do the memory copies */
733 0 : if (PREDICT_FALSE (vhost_user_input_copy (vui, cpu->copy, copy_len,
734 : &map_hint)))
735 : {
736 0 : vlib_error_count (vm, node->node_index,
737 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
738 : }
739 :
740 : /* give buffers back to driver */
741 0 : CLIB_MEMORY_STORE_BARRIER ();
742 0 : txvq->used->idx = txvq->last_used_idx;
743 0 : vhost_user_log_dirty_ring (vui, txvq, idx);
744 :
745 : /* interrupt (call) handling */
746 0 : if ((txvq->callfd_idx != ~0) &&
747 0 : !(txvq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
748 : {
749 0 : txvq->n_since_last_int += n_rx_packets;
750 :
751 0 : if (txvq->n_since_last_int > vum->coalesce_frames)
752 0 : vhost_user_send_call (vm, vui, txvq);
753 : }
754 :
755 : /* increase rx counters */
756 0 : vlib_increment_combined_counter
757 : (vnet_main.interface_main.combined_sw_if_counters
758 : + VNET_INTERFACE_COUNTER_RX, vm->thread_index, vui->sw_if_index,
759 : n_rx_packets, n_rx_bytes);
760 :
761 0 : vnet_device_increment_rx_packets (vm->thread_index, n_rx_packets);
762 :
763 161953 : done:
764 161953 : return n_rx_packets;
765 : }
766 :
767 : static_always_inline void
768 0 : vhost_user_mark_desc_consumed (vhost_user_intf_t * vui,
769 : vhost_user_vring_t * txvq, u16 desc_head,
770 : u16 n_descs_processed)
771 : {
772 0 : vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc;
773 : u16 desc_idx;
774 0 : u16 mask = txvq->qsz_mask;
775 :
776 0 : for (desc_idx = 0; desc_idx < n_descs_processed; desc_idx++)
777 : {
778 0 : if (txvq->used_wrap_counter)
779 0 : desc_table[(desc_head + desc_idx) & mask].flags |=
780 : (VRING_DESC_F_AVAIL | VRING_DESC_F_USED);
781 : else
782 0 : desc_table[(desc_head + desc_idx) & mask].flags &=
783 : ~(VRING_DESC_F_AVAIL | VRING_DESC_F_USED);
784 0 : vhost_user_advance_last_used_idx (txvq);
785 : }
786 0 : }
787 :
788 : static_always_inline void
789 0 : vhost_user_rx_trace_packed (vhost_trace_t * t, vhost_user_intf_t * vui,
790 : u16 qid, vhost_user_vring_t * txvq,
791 : u16 desc_current)
792 : {
793 0 : vhost_user_main_t *vum = &vhost_user_main;
794 : vnet_virtio_vring_packed_desc_t *hdr_desc;
795 : vnet_virtio_net_hdr_mrg_rxbuf_t *hdr;
796 0 : u32 hint = 0;
797 :
798 0 : clib_memset (t, 0, sizeof (*t));
799 0 : t->device_index = vui - vum->vhost_user_interfaces;
800 0 : t->qid = qid;
801 :
802 0 : hdr_desc = &txvq->packed_desc[desc_current];
803 0 : if (txvq->packed_desc[desc_current].flags & VRING_DESC_F_INDIRECT)
804 : {
805 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT;
806 : /* Header is the first here */
807 0 : hdr_desc = map_guest_mem (vui, txvq->packed_desc[desc_current].addr,
808 : &hint);
809 : }
810 0 : if (txvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT)
811 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED;
812 :
813 0 : if (!(txvq->packed_desc[desc_current].flags & VRING_DESC_F_NEXT) &&
814 0 : !(txvq->packed_desc[desc_current].flags & VRING_DESC_F_INDIRECT))
815 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC;
816 :
817 0 : t->first_desc_len = hdr_desc ? hdr_desc->len : 0;
818 :
819 0 : if (!hdr_desc || !(hdr = map_guest_mem (vui, hdr_desc->addr, &hint)))
820 0 : t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_MAP_ERROR;
821 : else
822 : {
823 0 : u32 len = vui->virtio_net_hdr_sz;
824 0 : clib_memcpy_fast (&t->hdr, hdr,
825 0 : len > hdr_desc->len ? hdr_desc->len : len);
826 : }
827 0 : }
828 :
829 : static_always_inline u32
830 0 : vhost_user_rx_discard_packet_packed (vlib_main_t * vm,
831 : vhost_user_intf_t * vui,
832 : vhost_user_vring_t * txvq,
833 : u32 discard_max)
834 : {
835 0 : u32 discarded_packets = 0;
836 0 : u16 mask = txvq->qsz_mask;
837 : u16 desc_current, desc_head;
838 :
839 0 : desc_head = desc_current = txvq->last_used_idx & mask;
840 :
841 : /*
842 : * On the RX side, each packet corresponds to one descriptor
843 : * (it is the same whether it is a shallow descriptor, chained, or indirect).
844 : * Therefore, discarding a packet is like discarding a descriptor.
845 : */
846 0 : while ((discarded_packets != discard_max) &&
847 0 : vhost_user_packed_desc_available (txvq, desc_current))
848 : {
849 0 : vhost_user_advance_last_avail_idx (txvq);
850 0 : discarded_packets++;
851 0 : desc_current = (desc_current + 1) & mask;
852 : }
853 :
854 0 : if (PREDICT_TRUE (discarded_packets))
855 0 : vhost_user_mark_desc_consumed (vui, txvq, desc_head, discarded_packets);
856 0 : return (discarded_packets);
857 : }
858 :
859 : static_always_inline u32
860 0 : vhost_user_input_copy_packed (vhost_user_intf_t * vui, vhost_copy_t * cpy,
861 : u16 copy_len, u32 * map_hint)
862 : {
863 : void *src0, *src1, *src2, *src3, *src4, *src5, *src6, *src7;
864 : u8 bad;
865 0 : u32 rc = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR;
866 :
867 0 : if (PREDICT_TRUE (copy_len >= 8))
868 : {
869 0 : src4 = map_guest_mem (vui, cpy[0].src, map_hint);
870 0 : src5 = map_guest_mem (vui, cpy[1].src, map_hint);
871 0 : src6 = map_guest_mem (vui, cpy[2].src, map_hint);
872 0 : src7 = map_guest_mem (vui, cpy[3].src, map_hint);
873 0 : bad = (src4 == 0) + (src5 == 0) + (src6 == 0) + (src7 == 0);
874 0 : if (PREDICT_FALSE (bad))
875 0 : goto one_by_one;
876 0 : clib_prefetch_load (src4);
877 0 : clib_prefetch_load (src5);
878 0 : clib_prefetch_load (src6);
879 0 : clib_prefetch_load (src7);
880 :
881 0 : while (PREDICT_TRUE (copy_len >= 8))
882 : {
883 0 : src0 = src4;
884 0 : src1 = src5;
885 0 : src2 = src6;
886 0 : src3 = src7;
887 :
888 0 : src4 = map_guest_mem (vui, cpy[4].src, map_hint);
889 0 : src5 = map_guest_mem (vui, cpy[5].src, map_hint);
890 0 : src6 = map_guest_mem (vui, cpy[6].src, map_hint);
891 0 : src7 = map_guest_mem (vui, cpy[7].src, map_hint);
892 0 : bad = (src4 == 0) + (src5 == 0) + (src6 == 0) + (src7 == 0);
893 0 : if (PREDICT_FALSE (bad))
894 0 : break;
895 :
896 0 : clib_prefetch_load (src4);
897 0 : clib_prefetch_load (src5);
898 0 : clib_prefetch_load (src6);
899 0 : clib_prefetch_load (src7);
900 :
901 0 : clib_memcpy_fast ((void *) cpy[0].dst, src0, cpy[0].len);
902 0 : clib_memcpy_fast ((void *) cpy[1].dst, src1, cpy[1].len);
903 0 : clib_memcpy_fast ((void *) cpy[2].dst, src2, cpy[2].len);
904 0 : clib_memcpy_fast ((void *) cpy[3].dst, src3, cpy[3].len);
905 0 : copy_len -= 4;
906 0 : cpy += 4;
907 : }
908 : }
909 :
910 0 : one_by_one:
911 0 : while (copy_len)
912 : {
913 0 : if (PREDICT_FALSE (!(src0 = map_guest_mem (vui, cpy->src, map_hint))))
914 : {
915 0 : rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL;
916 0 : break;
917 : }
918 0 : clib_memcpy_fast ((void *) cpy->dst, src0, cpy->len);
919 0 : copy_len -= 1;
920 0 : cpy += 1;
921 : }
922 0 : return rc;
923 : }
924 :
925 : static_always_inline u32
926 0 : vhost_user_do_offload (vhost_user_intf_t *vui,
927 : vnet_virtio_vring_packed_desc_t *desc_table,
928 : u16 desc_current, u16 mask, vlib_buffer_t *b_head,
929 : u32 *map_hint)
930 : {
931 0 : u32 rc = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR;
932 : vnet_virtio_net_hdr_mrg_rxbuf_t *hdr;
933 : u8 *b_data;
934 0 : u32 desc_data_offset = vui->virtio_net_hdr_sz;
935 :
936 0 : hdr = map_guest_mem (vui, desc_table[desc_current].addr, map_hint);
937 0 : if (PREDICT_FALSE (hdr == 0))
938 0 : rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL;
939 0 : else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
940 : {
941 0 : if (desc_data_offset == desc_table[desc_current].len)
942 : {
943 0 : desc_current = (desc_current + 1) & mask;
944 : b_data =
945 0 : map_guest_mem (vui, desc_table[desc_current].addr, map_hint);
946 0 : if (PREDICT_FALSE (b_data == 0))
947 0 : rc = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL;
948 : else
949 0 : vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr);
950 : }
951 : else
952 : {
953 0 : b_data = (u8 *) hdr + desc_data_offset;
954 0 : vhost_user_handle_rx_offload (b_head, b_data, &hdr->hdr);
955 : }
956 : }
957 :
958 0 : return rc;
959 : }
960 :
961 : static_always_inline u32
962 0 : vhost_user_compute_buffers_required (u32 desc_len, u32 buffer_data_size)
963 : {
964 : div_t result;
965 : u32 buffers_required;
966 :
967 0 : if (PREDICT_TRUE (buffer_data_size == 2048))
968 : {
969 0 : buffers_required = desc_len >> 11;
970 0 : if ((desc_len & 2047) != 0)
971 0 : buffers_required++;
972 0 : return (buffers_required);
973 : }
974 :
975 0 : if (desc_len < buffer_data_size)
976 0 : return 1;
977 :
978 0 : result = div (desc_len, buffer_data_size);
979 0 : if (result.rem)
980 0 : buffers_required = result.quot + 1;
981 : else
982 0 : buffers_required = result.quot;
983 :
984 0 : return (buffers_required);
985 : }
986 :
987 : static_always_inline u32
988 0 : vhost_user_compute_indirect_desc_len (vhost_user_intf_t * vui,
989 : vhost_user_vring_t * txvq,
990 : u32 buffer_data_size, u16 desc_current,
991 : u32 * map_hint)
992 : {
993 0 : vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc;
994 0 : u32 desc_len = 0;
995 0 : u16 desc_data_offset = vui->virtio_net_hdr_sz;
996 0 : u16 desc_idx = desc_current;
997 : u32 n_descs;
998 :
999 0 : n_descs = desc_table[desc_idx].len >> 4;
1000 0 : desc_table = map_guest_mem (vui, desc_table[desc_idx].addr, map_hint);
1001 0 : if (PREDICT_FALSE (desc_table == 0))
1002 0 : return 0;
1003 :
1004 0 : for (desc_idx = 0; desc_idx < n_descs; desc_idx++)
1005 0 : desc_len += desc_table[desc_idx].len;
1006 :
1007 0 : if (PREDICT_TRUE (desc_len > desc_data_offset))
1008 0 : desc_len -= desc_data_offset;
1009 :
1010 0 : return vhost_user_compute_buffers_required (desc_len, buffer_data_size);
1011 : }
1012 :
1013 : static_always_inline u32
1014 0 : vhost_user_compute_chained_desc_len (vhost_user_intf_t * vui,
1015 : vhost_user_vring_t * txvq,
1016 : u32 buffer_data_size, u16 * current,
1017 : u16 * n_left)
1018 : {
1019 0 : vnet_virtio_vring_packed_desc_t *desc_table = txvq->packed_desc;
1020 0 : u32 desc_len = 0;
1021 0 : u16 mask = txvq->qsz_mask;
1022 :
1023 0 : while (desc_table[*current].flags & VRING_DESC_F_NEXT)
1024 : {
1025 0 : desc_len += desc_table[*current].len;
1026 0 : (*n_left)++;
1027 0 : *current = (*current + 1) & mask;
1028 0 : vhost_user_advance_last_avail_idx (txvq);
1029 : }
1030 0 : desc_len += desc_table[*current].len;
1031 0 : (*n_left)++;
1032 0 : *current = (*current + 1) & mask;
1033 0 : vhost_user_advance_last_avail_idx (txvq);
1034 :
1035 0 : if (PREDICT_TRUE (desc_len > vui->virtio_net_hdr_sz))
1036 0 : desc_len -= vui->virtio_net_hdr_sz;
1037 :
1038 0 : return vhost_user_compute_buffers_required (desc_len, buffer_data_size);
1039 : }
1040 :
1041 : static_always_inline void
1042 0 : vhost_user_assemble_packet (vnet_virtio_vring_packed_desc_t *desc_table,
1043 : u16 *desc_idx, vlib_buffer_t *b_head,
1044 : vlib_buffer_t **b_current, u32 **next,
1045 : vlib_buffer_t ***b, u32 *bi_current,
1046 : vhost_cpu_t *cpu, u16 *copy_len, u32 *buffers_used,
1047 : u32 buffers_required, u32 *desc_data_offset,
1048 : u32 buffer_data_size, u16 mask)
1049 : {
1050 : u32 desc_data_l;
1051 :
1052 0 : while (*desc_data_offset < desc_table[*desc_idx].len)
1053 : {
1054 : /* Get more output if necessary. Or end of packet. */
1055 0 : if (PREDICT_FALSE ((*b_current)->current_length == buffer_data_size))
1056 : {
1057 : /* Get next output */
1058 0 : u32 bi_next = **next;
1059 0 : (*next)++;
1060 0 : (*b_current)->next_buffer = bi_next;
1061 0 : (*b_current)->flags |= VLIB_BUFFER_NEXT_PRESENT;
1062 0 : *bi_current = bi_next;
1063 0 : *b_current = **b;
1064 0 : (*b)++;
1065 0 : (*buffers_used)++;
1066 0 : ASSERT (*buffers_used <= buffers_required);
1067 : }
1068 :
1069 : /* Prepare a copy order executed later for the data */
1070 0 : ASSERT (*copy_len < VHOST_USER_COPY_ARRAY_N);
1071 0 : vhost_copy_t *cpy = &cpu->copy[*copy_len];
1072 0 : (*copy_len)++;
1073 0 : desc_data_l = desc_table[*desc_idx].len - *desc_data_offset;
1074 0 : cpy->len = buffer_data_size - (*b_current)->current_length;
1075 0 : cpy->len = (cpy->len > desc_data_l) ? desc_data_l : cpy->len;
1076 0 : cpy->dst = (uword) (vlib_buffer_get_current (*b_current) +
1077 0 : (*b_current)->current_length);
1078 0 : cpy->src = desc_table[*desc_idx].addr + *desc_data_offset;
1079 :
1080 0 : *desc_data_offset += cpy->len;
1081 :
1082 0 : (*b_current)->current_length += cpy->len;
1083 0 : b_head->total_length_not_including_first_buffer += cpy->len;
1084 : }
1085 0 : *desc_idx = (*desc_idx + 1) & mask;;
1086 0 : *desc_data_offset = 0;
1087 0 : }
1088 :
1089 : static_always_inline u32
1090 0 : vhost_user_if_input_packed (vlib_main_t *vm, vhost_user_main_t *vum,
1091 : vhost_user_intf_t *vui, u16 qid,
1092 : vlib_node_runtime_t *node, u8 enable_csum)
1093 : {
1094 0 : vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)];
1095 0 : vnet_feature_main_t *fm = &feature_main;
1096 0 : u8 feature_arc_idx = fm->device_input_feature_arc_index;
1097 0 : u16 n_rx_packets = 0;
1098 0 : u32 n_rx_bytes = 0;
1099 0 : u16 n_left = 0;
1100 0 : u32 buffers_required = 0;
1101 : u32 n_left_to_next, *to_next;
1102 0 : u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
1103 0 : u32 n_trace = vlib_get_trace_count (vm, node);
1104 0 : u32 buffer_data_size = vlib_buffer_get_default_data_size (vm);
1105 0 : u32 map_hint = 0;
1106 0 : vhost_cpu_t *cpu = &vum->cpus[vm->thread_index];
1107 0 : u16 copy_len = 0;
1108 0 : u32 current_config_index = ~0;
1109 0 : u16 mask = txvq->qsz_mask;
1110 : u16 desc_current, desc_head, last_used_idx;
1111 0 : vnet_virtio_vring_packed_desc_t *desc_table = 0;
1112 0 : u32 n_descs_processed = 0;
1113 : u32 rv;
1114 : vlib_buffer_t **b;
1115 : u32 *next;
1116 0 : u32 buffers_used = 0;
1117 : u16 current, n_descs_to_process;
1118 :
1119 : /* The descriptor table is not ready yet */
1120 0 : if (PREDICT_FALSE (txvq->packed_desc == 0))
1121 0 : goto done;
1122 :
1123 : /* do we have pending interrupts ? */
1124 0 : vhost_user_vring_t *rxvq = &vui->vrings[VHOST_VRING_IDX_RX (qid)];
1125 0 : vhost_user_input_do_interrupt (vm, vui, txvq, rxvq);
1126 :
1127 : /*
1128 : * For adaptive mode, it is optimized to reduce interrupts.
1129 : * If the scheduler switches the input node to polling due
1130 : * to burst of traffic, we tell the driver no interrupt.
1131 : * When the traffic subsides, the scheduler switches the node back to
1132 : * interrupt mode. We must tell the driver we want interrupt.
1133 : */
1134 0 : if (PREDICT_FALSE (txvq->mode == VNET_HW_IF_RX_MODE_ADAPTIVE))
1135 : {
1136 0 : if ((node->flags &
1137 0 : VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE) ||
1138 0 : !(node->flags &
1139 : VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE))
1140 : /* Tell driver we want notification */
1141 0 : txvq->used_event->flags = 0;
1142 : else
1143 : /* Tell driver we don't want notification */
1144 0 : txvq->used_event->flags = VRING_EVENT_F_DISABLE;
1145 : }
1146 :
1147 0 : last_used_idx = txvq->last_used_idx & mask;
1148 0 : desc_head = desc_current = last_used_idx;
1149 :
1150 0 : if (vhost_user_packed_desc_available (txvq, desc_current) == 0)
1151 0 : goto done;
1152 :
1153 0 : if (PREDICT_FALSE (!vui->admin_up || !vui->is_ready || !(txvq->enabled)))
1154 : {
1155 : /*
1156 : * Discard input packet if interface is admin down or vring is not
1157 : * enabled.
1158 : * "For example, for a networking device, in the disabled state
1159 : * client must not supply any new RX packets, but must process
1160 : * and discard any TX packets."
1161 : */
1162 0 : rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq,
1163 : VHOST_USER_DOWN_DISCARD_COUNT);
1164 0 : vlib_error_count (vm, vhost_user_input_node.index,
1165 : VHOST_USER_INPUT_FUNC_ERROR_NOT_READY, rv);
1166 0 : goto done;
1167 : }
1168 :
1169 0 : vhost_user_input_setup_frame (vm, node, vui, ¤t_config_index,
1170 : &next_index, &to_next, &n_left_to_next);
1171 :
1172 : /*
1173 : * Compute n_left and total buffers needed
1174 : */
1175 0 : desc_table = txvq->packed_desc;
1176 0 : current = desc_current;
1177 0 : while (vhost_user_packed_desc_available (txvq, current) &&
1178 0 : (n_left < VLIB_FRAME_SIZE))
1179 : {
1180 0 : if (desc_table[current].flags & VRING_DESC_F_INDIRECT)
1181 : {
1182 0 : buffers_required +=
1183 0 : vhost_user_compute_indirect_desc_len (vui, txvq, buffer_data_size,
1184 : current, &map_hint);
1185 0 : n_left++;
1186 0 : current = (current + 1) & mask;
1187 0 : vhost_user_advance_last_avail_idx (txvq);
1188 : }
1189 : else
1190 : {
1191 0 : buffers_required +=
1192 0 : vhost_user_compute_chained_desc_len (vui, txvq, buffer_data_size,
1193 : ¤t, &n_left);
1194 : }
1195 : }
1196 :
1197 : /* Something is broken if we need more than 10000 buffers */
1198 0 : if (PREDICT_FALSE ((buffers_required == 0) || (buffers_required > 10000)))
1199 : {
1200 0 : rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq, n_left);
1201 0 : vlib_error_count (vm, vhost_user_input_node.index,
1202 : VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, rv);
1203 0 : goto done;
1204 : }
1205 :
1206 0 : vec_validate (cpu->to_next_list, buffers_required);
1207 0 : rv = vlib_buffer_alloc (vm, cpu->to_next_list, buffers_required);
1208 0 : if (PREDICT_FALSE (rv != buffers_required))
1209 : {
1210 0 : vlib_buffer_free (vm, cpu->to_next_list, rv);
1211 0 : rv = vhost_user_rx_discard_packet_packed (vm, vui, txvq, n_left);
1212 0 : vlib_error_count (vm, vhost_user_input_node.index,
1213 : VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, rv);
1214 0 : goto done;
1215 : }
1216 :
1217 0 : next = cpu->to_next_list;
1218 0 : vec_validate (cpu->rx_buffers_pdesc, buffers_required);
1219 0 : vlib_get_buffers (vm, next, cpu->rx_buffers_pdesc, buffers_required);
1220 0 : b = cpu->rx_buffers_pdesc;
1221 0 : n_descs_processed = n_left;
1222 :
1223 0 : while (n_left)
1224 : {
1225 : vlib_buffer_t *b_head, *b_current;
1226 : u32 bi_current;
1227 : u32 desc_data_offset;
1228 0 : u16 desc_idx = desc_current;
1229 : u32 n_descs;
1230 :
1231 0 : desc_table = txvq->packed_desc;
1232 0 : to_next[0] = bi_current = next[0];
1233 0 : b_head = b_current = b[0];
1234 0 : b++;
1235 0 : buffers_used++;
1236 0 : ASSERT (buffers_used <= buffers_required);
1237 0 : to_next++;
1238 0 : next++;
1239 0 : n_left_to_next--;
1240 :
1241 : /* The buffer should already be initialized */
1242 0 : b_head->total_length_not_including_first_buffer = 0;
1243 0 : b_head->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
1244 0 : desc_data_offset = vui->virtio_net_hdr_sz;
1245 0 : n_descs_to_process = 1;
1246 :
1247 0 : if (desc_table[desc_idx].flags & VRING_DESC_F_INDIRECT)
1248 : {
1249 0 : n_descs = desc_table[desc_idx].len >> 4;
1250 0 : desc_table = map_guest_mem (vui, desc_table[desc_idx].addr,
1251 : &map_hint);
1252 0 : desc_idx = 0;
1253 0 : if (PREDICT_FALSE (desc_table == 0) ||
1254 0 : (enable_csum &&
1255 0 : (PREDICT_FALSE
1256 : (vhost_user_do_offload
1257 : (vui, desc_table, desc_idx, mask, b_head,
1258 : &map_hint) != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR))))
1259 : {
1260 0 : vlib_error_count (vm, node->node_index,
1261 : VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL, 1);
1262 0 : to_next--;
1263 0 : next--;
1264 0 : n_left_to_next++;
1265 0 : buffers_used--;
1266 0 : b--;
1267 0 : goto out;
1268 : }
1269 0 : while (n_descs)
1270 : {
1271 0 : vhost_user_assemble_packet (desc_table, &desc_idx, b_head,
1272 : &b_current, &next, &b, &bi_current,
1273 : cpu, ©_len, &buffers_used,
1274 : buffers_required, &desc_data_offset,
1275 : buffer_data_size, mask);
1276 0 : n_descs--;
1277 : }
1278 : }
1279 : else
1280 : {
1281 0 : if (enable_csum)
1282 : {
1283 0 : rv = vhost_user_do_offload (vui, desc_table, desc_idx, mask,
1284 : b_head, &map_hint);
1285 0 : if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR))
1286 : {
1287 0 : vlib_error_count (vm, node->node_index, rv, 1);
1288 0 : to_next--;
1289 0 : next--;
1290 0 : n_left_to_next++;
1291 0 : buffers_used--;
1292 0 : b--;
1293 0 : goto out;
1294 : }
1295 : }
1296 : /*
1297 : * For chained descriptor, we process all chains in a single while
1298 : * loop. So count how many descriptors in the chain.
1299 : */
1300 0 : n_descs_to_process = 1;
1301 0 : while (desc_table[desc_idx].flags & VRING_DESC_F_NEXT)
1302 : {
1303 0 : vhost_user_assemble_packet (desc_table, &desc_idx, b_head,
1304 : &b_current, &next, &b, &bi_current,
1305 : cpu, ©_len, &buffers_used,
1306 : buffers_required, &desc_data_offset,
1307 : buffer_data_size, mask);
1308 0 : n_descs_to_process++;
1309 : }
1310 0 : vhost_user_assemble_packet (desc_table, &desc_idx, b_head,
1311 : &b_current, &next, &b, &bi_current,
1312 : cpu, ©_len, &buffers_used,
1313 : buffers_required, &desc_data_offset,
1314 : buffer_data_size, mask);
1315 : }
1316 :
1317 0 : n_rx_bytes += b_head->total_length_not_including_first_buffer;
1318 0 : n_rx_packets++;
1319 :
1320 0 : b_head->total_length_not_including_first_buffer -=
1321 0 : b_head->current_length;
1322 :
1323 0 : vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index;
1324 0 : vnet_buffer (b_head)->sw_if_index[VLIB_TX] = ~0;
1325 0 : b_head->error = 0;
1326 :
1327 0 : if (current_config_index != ~0)
1328 : {
1329 0 : b_head->current_config_index = current_config_index;
1330 0 : vnet_buffer (b_head)->feature_arc_index = feature_arc_idx;
1331 : }
1332 :
1333 0 : out:
1334 0 : ASSERT (n_left >= n_descs_to_process);
1335 0 : n_left -= n_descs_to_process;
1336 :
1337 : /* advance to next descrptor */
1338 0 : desc_current = (desc_current + n_descs_to_process) & mask;
1339 :
1340 : /*
1341 : * Although separating memory copies from virtio ring parsing
1342 : * is beneficial, we can offer to perform the copies from time
1343 : * to time in order to free some space in the ring.
1344 : */
1345 0 : if (PREDICT_FALSE (copy_len >= VHOST_USER_RX_COPY_THRESHOLD))
1346 : {
1347 0 : rv = vhost_user_input_copy_packed (vui, cpu->copy, copy_len,
1348 : &map_hint);
1349 0 : if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR))
1350 0 : vlib_error_count (vm, node->node_index, rv, 1);
1351 0 : copy_len = 0;
1352 : }
1353 : }
1354 0 : vlib_put_next_frame (vm, node, next_index, n_left_to_next);
1355 :
1356 : /* Do the memory copies */
1357 0 : rv = vhost_user_input_copy_packed (vui, cpu->copy, copy_len, &map_hint);
1358 0 : if (PREDICT_FALSE (rv != VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR))
1359 0 : vlib_error_count (vm, node->node_index, rv, 1);
1360 :
1361 : /* Must do the tracing before giving buffers back to driver */
1362 0 : if (PREDICT_FALSE (n_trace))
1363 : {
1364 0 : u32 left = n_rx_packets;
1365 :
1366 0 : b = cpu->rx_buffers_pdesc;
1367 0 : while (n_trace && left)
1368 : {
1369 0 : if (PREDICT_TRUE
1370 : (vlib_trace_buffer
1371 : (vm, node, next_index, b[0], /* follow_chain */ 0)))
1372 : {
1373 : vhost_trace_t *t0;
1374 0 : t0 = vlib_add_trace (vm, node, b[0], sizeof (t0[0]));
1375 0 : vhost_user_rx_trace_packed (t0, vui, qid, txvq, last_used_idx);
1376 0 : last_used_idx = (last_used_idx + 1) & mask;
1377 0 : n_trace--;
1378 0 : vlib_set_trace_count (vm, node, n_trace);
1379 : }
1380 0 : left--;
1381 0 : b++;
1382 : }
1383 : }
1384 :
1385 : /*
1386 : * Give buffers back to driver.
1387 : */
1388 0 : vhost_user_mark_desc_consumed (vui, txvq, desc_head, n_descs_processed);
1389 :
1390 : /* interrupt (call) handling */
1391 0 : if ((txvq->callfd_idx != ~0) &&
1392 0 : (txvq->avail_event->flags != VRING_EVENT_F_DISABLE))
1393 : {
1394 0 : txvq->n_since_last_int += n_rx_packets;
1395 0 : if (txvq->n_since_last_int > vum->coalesce_frames)
1396 0 : vhost_user_send_call (vm, vui, txvq);
1397 : }
1398 :
1399 : /* increase rx counters */
1400 0 : vlib_increment_combined_counter
1401 : (vnet_main.interface_main.combined_sw_if_counters
1402 : + VNET_INTERFACE_COUNTER_RX, vm->thread_index, vui->sw_if_index,
1403 : n_rx_packets, n_rx_bytes);
1404 :
1405 0 : vnet_device_increment_rx_packets (vm->thread_index, n_rx_packets);
1406 :
1407 0 : if (PREDICT_FALSE (buffers_used < buffers_required))
1408 0 : vlib_buffer_free (vm, next, buffers_required - buffers_used);
1409 :
1410 0 : done:
1411 0 : return n_rx_packets;
1412 : }
1413 :
1414 159128 : VLIB_NODE_FN (vhost_user_input_node) (vlib_main_t * vm,
1415 : vlib_node_runtime_t * node,
1416 : vlib_frame_t * frame)
1417 : {
1418 156828 : vhost_user_main_t *vum = &vhost_user_main;
1419 156828 : uword n_rx_packets = 0;
1420 : vhost_user_intf_t *vui;
1421 156828 : vnet_hw_if_rxq_poll_vector_t *pv = vnet_hw_if_get_rxq_poll_vector (vm, node);
1422 : vnet_hw_if_rxq_poll_vector_t *pve;
1423 :
1424 318781 : vec_foreach (pve, pv)
1425 : {
1426 161953 : vui = pool_elt_at_index (vum->vhost_user_interfaces, pve->dev_instance);
1427 161953 : if (vhost_user_is_packed_ring_supported (vui))
1428 : {
1429 0 : if (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_CSUM))
1430 0 : n_rx_packets += vhost_user_if_input_packed (
1431 0 : vm, vum, vui, pve->queue_id, node, 1);
1432 : else
1433 0 : n_rx_packets += vhost_user_if_input_packed (
1434 0 : vm, vum, vui, pve->queue_id, node, 0);
1435 : }
1436 : else
1437 : {
1438 161953 : if (vui->features & VIRTIO_FEATURE (VIRTIO_NET_F_CSUM))
1439 0 : n_rx_packets +=
1440 0 : vhost_user_if_input (vm, vum, vui, pve->queue_id, node, 1);
1441 : else
1442 161953 : n_rx_packets +=
1443 161953 : vhost_user_if_input (vm, vum, vui, pve->queue_id, node, 0);
1444 : }
1445 : }
1446 :
1447 156828 : return n_rx_packets;
1448 : }
1449 :
1450 : /* *INDENT-OFF* */
1451 17855 : VLIB_REGISTER_NODE (vhost_user_input_node) = {
1452 : .type = VLIB_NODE_TYPE_INPUT,
1453 : .name = "vhost-user-input",
1454 : .sibling_of = "device-input",
1455 : .flags = VLIB_NODE_FLAG_TRACE_SUPPORTED,
1456 :
1457 : /* Will be enabled if/when hardware is detected. */
1458 : .state = VLIB_NODE_STATE_DISABLED,
1459 :
1460 : .format_buffer = format_ethernet_header_with_length,
1461 : .format_trace = format_vhost_trace,
1462 :
1463 : .n_errors = VHOST_USER_INPUT_FUNC_N_ERROR,
1464 : .error_strings = vhost_user_input_func_error_strings,
1465 : };
1466 : /* *INDENT-ON* */
1467 :
1468 : /*
1469 : * fd.io coding-style-patch-verification: ON
1470 : *
1471 : * Local Variables:
1472 : * eval: (c-set-style "gnu")
1473 : * End:
1474 : */
|
