Mohsin Kazmi | e7cde31 | 2018-06-26 17:20:11 +0200 | [diff] [blame] | 1 | /* |
| 2 | *------------------------------------------------------------------ |
| 3 | * vhost-user-output |
| 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 | * On the transmit side, we keep processing the buffers from vlib in the while |
| 47 | * loop and prepare the copy order to be executed later. However, the static |
| 48 | * array which we keep the copy order is limited to VHOST_USER_COPY_ARRAY_N |
| 49 | * entries. In order to not corrupt memory, we have to do the copy when the |
| 50 | * static array reaches the copy threshold. We subtract 40 in case the code |
| 51 | * goes into the inner loop for a maximum of 64k frames which may require |
| 52 | * more array entries. |
| 53 | */ |
| 54 | #define VHOST_USER_TX_COPY_THRESHOLD (VHOST_USER_COPY_ARRAY_N - 40) |
| 55 | |
| 56 | vnet_device_class_t vhost_user_device_class; |
| 57 | |
| 58 | #define foreach_vhost_user_tx_func_error \ |
| 59 | _(NONE, "no error") \ |
| 60 | _(NOT_READY, "vhost vring not ready") \ |
| 61 | _(DOWN, "vhost interface is down") \ |
| 62 | _(PKT_DROP_NOBUF, "tx packet drops (no available descriptors)") \ |
| 63 | _(PKT_DROP_NOMRG, "tx packet drops (cannot merge descriptors)") \ |
| 64 | _(MMAP_FAIL, "mmap failure") \ |
| 65 | _(INDIRECT_OVERFLOW, "indirect descriptor table overflow") |
| 66 | |
| 67 | typedef enum |
| 68 | { |
| 69 | #define _(f,s) VHOST_USER_TX_FUNC_ERROR_##f, |
| 70 | foreach_vhost_user_tx_func_error |
| 71 | #undef _ |
| 72 | VHOST_USER_TX_FUNC_N_ERROR, |
| 73 | } vhost_user_tx_func_error_t; |
| 74 | |
| 75 | static __clib_unused char *vhost_user_tx_func_error_strings[] = { |
| 76 | #define _(n,s) s, |
| 77 | foreach_vhost_user_tx_func_error |
| 78 | #undef _ |
| 79 | }; |
| 80 | |
| 81 | static __clib_unused u8 * |
| 82 | format_vhost_user_interface_name (u8 * s, va_list * args) |
| 83 | { |
| 84 | u32 i = va_arg (*args, u32); |
| 85 | u32 show_dev_instance = ~0; |
| 86 | vhost_user_main_t *vum = &vhost_user_main; |
| 87 | |
| 88 | if (i < vec_len (vum->show_dev_instance_by_real_dev_instance)) |
| 89 | show_dev_instance = vum->show_dev_instance_by_real_dev_instance[i]; |
| 90 | |
| 91 | if (show_dev_instance != ~0) |
| 92 | i = show_dev_instance; |
| 93 | |
| 94 | s = format (s, "VirtualEthernet0/0/%d", i); |
| 95 | return s; |
| 96 | } |
| 97 | |
| 98 | static __clib_unused int |
| 99 | vhost_user_name_renumber (vnet_hw_interface_t * hi, u32 new_dev_instance) |
| 100 | { |
| 101 | // FIXME: check if the new dev instance is already used |
| 102 | vhost_user_main_t *vum = &vhost_user_main; |
| 103 | vec_validate_init_empty (vum->show_dev_instance_by_real_dev_instance, |
| 104 | hi->dev_instance, ~0); |
| 105 | |
| 106 | vum->show_dev_instance_by_real_dev_instance[hi->dev_instance] = |
| 107 | new_dev_instance; |
| 108 | |
| 109 | DBG_SOCK ("renumbered vhost-user interface dev_instance %d to %d", |
| 110 | hi->dev_instance, new_dev_instance); |
| 111 | |
| 112 | return 0; |
| 113 | } |
| 114 | |
| 115 | /** |
| 116 | * @brief Try once to lock the vring |
| 117 | * @return 0 on success, non-zero on failure. |
| 118 | */ |
| 119 | static_always_inline int |
| 120 | vhost_user_vring_try_lock (vhost_user_intf_t * vui, u32 qid) |
| 121 | { |
| 122 | return __sync_lock_test_and_set (vui->vring_locks[qid], 1); |
| 123 | } |
| 124 | |
| 125 | /** |
| 126 | * @brief Spin until the vring is successfully locked |
| 127 | */ |
| 128 | static_always_inline void |
| 129 | vhost_user_vring_lock (vhost_user_intf_t * vui, u32 qid) |
| 130 | { |
| 131 | while (vhost_user_vring_try_lock (vui, qid)) |
| 132 | ; |
| 133 | } |
| 134 | |
| 135 | /** |
| 136 | * @brief Unlock the vring lock |
| 137 | */ |
| 138 | static_always_inline void |
| 139 | vhost_user_vring_unlock (vhost_user_intf_t * vui, u32 qid) |
| 140 | { |
| 141 | *vui->vring_locks[qid] = 0; |
| 142 | } |
| 143 | |
| 144 | static_always_inline void |
| 145 | vhost_user_tx_trace (vhost_trace_t * t, |
| 146 | vhost_user_intf_t * vui, u16 qid, |
| 147 | vlib_buffer_t * b, vhost_user_vring_t * rxvq) |
| 148 | { |
| 149 | vhost_user_main_t *vum = &vhost_user_main; |
| 150 | u32 last_avail_idx = rxvq->last_avail_idx; |
| 151 | u32 desc_current = rxvq->avail->ring[last_avail_idx & rxvq->qsz_mask]; |
| 152 | vring_desc_t *hdr_desc = 0; |
| 153 | u32 hint = 0; |
| 154 | |
| 155 | memset (t, 0, sizeof (*t)); |
| 156 | t->device_index = vui - vum->vhost_user_interfaces; |
| 157 | t->qid = qid; |
| 158 | |
| 159 | hdr_desc = &rxvq->desc[desc_current]; |
| 160 | if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT) |
| 161 | { |
| 162 | t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_INDIRECT; |
| 163 | /* Header is the first here */ |
| 164 | hdr_desc = map_guest_mem (vui, rxvq->desc[desc_current].addr, &hint); |
| 165 | } |
| 166 | if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) |
| 167 | { |
| 168 | t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SIMPLE_CHAINED; |
| 169 | } |
| 170 | if (!(rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) && |
| 171 | !(rxvq->desc[desc_current].flags & VIRTQ_DESC_F_INDIRECT)) |
| 172 | { |
| 173 | t->virtio_ring_flags |= 1 << VIRTIO_TRACE_F_SINGLE_DESC; |
| 174 | } |
| 175 | |
| 176 | t->first_desc_len = hdr_desc ? hdr_desc->len : 0; |
| 177 | } |
| 178 | |
| 179 | static_always_inline u32 |
| 180 | vhost_user_tx_copy (vhost_user_intf_t * vui, vhost_copy_t * cpy, |
| 181 | u16 copy_len, u32 * map_hint) |
| 182 | { |
| 183 | void *dst0, *dst1, *dst2, *dst3; |
| 184 | if (PREDICT_TRUE (copy_len >= 4)) |
| 185 | { |
| 186 | if (PREDICT_FALSE (!(dst2 = map_guest_mem (vui, cpy[0].dst, map_hint)))) |
| 187 | return 1; |
| 188 | if (PREDICT_FALSE (!(dst3 = map_guest_mem (vui, cpy[1].dst, map_hint)))) |
| 189 | return 1; |
| 190 | while (PREDICT_TRUE (copy_len >= 4)) |
| 191 | { |
| 192 | dst0 = dst2; |
| 193 | dst1 = dst3; |
| 194 | |
| 195 | if (PREDICT_FALSE |
| 196 | (!(dst2 = map_guest_mem (vui, cpy[2].dst, map_hint)))) |
| 197 | return 1; |
| 198 | if (PREDICT_FALSE |
| 199 | (!(dst3 = map_guest_mem (vui, cpy[3].dst, map_hint)))) |
| 200 | return 1; |
| 201 | |
| 202 | CLIB_PREFETCH ((void *) cpy[2].src, 64, LOAD); |
| 203 | CLIB_PREFETCH ((void *) cpy[3].src, 64, LOAD); |
| 204 | |
| 205 | clib_memcpy (dst0, (void *) cpy[0].src, cpy[0].len); |
| 206 | clib_memcpy (dst1, (void *) cpy[1].src, cpy[1].len); |
| 207 | |
| 208 | vhost_user_log_dirty_pages_2 (vui, cpy[0].dst, cpy[0].len, 1); |
| 209 | vhost_user_log_dirty_pages_2 (vui, cpy[1].dst, cpy[1].len, 1); |
| 210 | copy_len -= 2; |
| 211 | cpy += 2; |
| 212 | } |
| 213 | } |
| 214 | while (copy_len) |
| 215 | { |
| 216 | if (PREDICT_FALSE (!(dst0 = map_guest_mem (vui, cpy->dst, map_hint)))) |
| 217 | return 1; |
| 218 | clib_memcpy (dst0, (void *) cpy->src, cpy->len); |
| 219 | vhost_user_log_dirty_pages_2 (vui, cpy->dst, cpy->len, 1); |
| 220 | copy_len -= 1; |
| 221 | cpy += 1; |
| 222 | } |
| 223 | return 0; |
| 224 | } |
| 225 | |
| 226 | |
| 227 | uword |
| 228 | CLIB_MULTIARCH_FN (vhost_user_tx) (vlib_main_t * vm, |
| 229 | vlib_node_runtime_t * node, |
| 230 | vlib_frame_t * frame) |
| 231 | { |
| 232 | u32 *buffers = vlib_frame_args (frame); |
| 233 | u32 n_left = frame->n_vectors; |
| 234 | vhost_user_main_t *vum = &vhost_user_main; |
| 235 | vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; |
| 236 | vhost_user_intf_t *vui = |
| 237 | pool_elt_at_index (vum->vhost_user_interfaces, rd->dev_instance); |
| 238 | u32 qid = ~0; |
| 239 | vhost_user_vring_t *rxvq; |
| 240 | u8 error; |
| 241 | u32 thread_index = vlib_get_thread_index (); |
| 242 | u32 map_hint = 0; |
| 243 | u8 retry = 8; |
| 244 | u16 copy_len; |
| 245 | u16 tx_headers_len; |
| 246 | |
| 247 | if (PREDICT_FALSE (!vui->admin_up)) |
| 248 | { |
| 249 | error = VHOST_USER_TX_FUNC_ERROR_DOWN; |
| 250 | goto done3; |
| 251 | } |
| 252 | |
| 253 | if (PREDICT_FALSE (!vui->is_up)) |
| 254 | { |
| 255 | error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; |
| 256 | goto done3; |
| 257 | } |
| 258 | |
| 259 | qid = |
| 260 | VHOST_VRING_IDX_RX (*vec_elt_at_index |
| 261 | (vui->per_cpu_tx_qid, thread_index)); |
| 262 | rxvq = &vui->vrings[qid]; |
| 263 | if (PREDICT_FALSE (vui->use_tx_spinlock)) |
| 264 | vhost_user_vring_lock (vui, qid); |
| 265 | |
| 266 | retry: |
| 267 | error = VHOST_USER_TX_FUNC_ERROR_NONE; |
| 268 | tx_headers_len = 0; |
| 269 | copy_len = 0; |
| 270 | while (n_left > 0) |
| 271 | { |
| 272 | vlib_buffer_t *b0, *current_b0; |
| 273 | u16 desc_head, desc_index, desc_len; |
| 274 | vring_desc_t *desc_table; |
| 275 | uword buffer_map_addr; |
| 276 | u32 buffer_len; |
| 277 | u16 bytes_left; |
| 278 | |
| 279 | if (PREDICT_TRUE (n_left > 1)) |
| 280 | vlib_prefetch_buffer_with_index (vm, buffers[1], LOAD); |
| 281 | |
| 282 | b0 = vlib_get_buffer (vm, buffers[0]); |
| 283 | |
| 284 | if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) |
| 285 | { |
| 286 | vum->cpus[thread_index].current_trace = |
| 287 | vlib_add_trace (vm, node, b0, |
| 288 | sizeof (*vum->cpus[thread_index].current_trace)); |
| 289 | vhost_user_tx_trace (vum->cpus[thread_index].current_trace, |
| 290 | vui, qid / 2, b0, rxvq); |
| 291 | } |
| 292 | |
| 293 | if (PREDICT_FALSE (rxvq->last_avail_idx == rxvq->avail->idx)) |
| 294 | { |
| 295 | error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; |
| 296 | goto done; |
| 297 | } |
| 298 | |
| 299 | desc_table = rxvq->desc; |
| 300 | desc_head = desc_index = |
| 301 | rxvq->avail->ring[rxvq->last_avail_idx & rxvq->qsz_mask]; |
| 302 | |
| 303 | /* Go deeper in case of indirect descriptor |
| 304 | * I don't know of any driver providing indirect for RX. */ |
| 305 | if (PREDICT_FALSE (rxvq->desc[desc_head].flags & VIRTQ_DESC_F_INDIRECT)) |
| 306 | { |
| 307 | if (PREDICT_FALSE |
| 308 | (rxvq->desc[desc_head].len < sizeof (vring_desc_t))) |
| 309 | { |
| 310 | error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; |
| 311 | goto done; |
| 312 | } |
| 313 | if (PREDICT_FALSE |
| 314 | (!(desc_table = |
| 315 | map_guest_mem (vui, rxvq->desc[desc_index].addr, |
| 316 | &map_hint)))) |
| 317 | { |
| 318 | error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; |
| 319 | goto done; |
| 320 | } |
| 321 | desc_index = 0; |
| 322 | } |
| 323 | |
| 324 | desc_len = vui->virtio_net_hdr_sz; |
| 325 | buffer_map_addr = desc_table[desc_index].addr; |
| 326 | buffer_len = desc_table[desc_index].len; |
| 327 | |
| 328 | { |
| 329 | // Get a header from the header array |
| 330 | virtio_net_hdr_mrg_rxbuf_t *hdr = |
| 331 | &vum->cpus[thread_index].tx_headers[tx_headers_len]; |
| 332 | tx_headers_len++; |
| 333 | hdr->hdr.flags = 0; |
| 334 | hdr->hdr.gso_type = 0; |
| 335 | hdr->num_buffers = 1; //This is local, no need to check |
| 336 | |
| 337 | // Prepare a copy order executed later for the header |
| 338 | vhost_copy_t *cpy = &vum->cpus[thread_index].copy[copy_len]; |
| 339 | copy_len++; |
| 340 | cpy->len = vui->virtio_net_hdr_sz; |
| 341 | cpy->dst = buffer_map_addr; |
| 342 | cpy->src = (uword) hdr; |
| 343 | } |
| 344 | |
| 345 | buffer_map_addr += vui->virtio_net_hdr_sz; |
| 346 | buffer_len -= vui->virtio_net_hdr_sz; |
| 347 | bytes_left = b0->current_length; |
| 348 | current_b0 = b0; |
| 349 | while (1) |
| 350 | { |
| 351 | if (buffer_len == 0) |
| 352 | { //Get new output |
| 353 | if (desc_table[desc_index].flags & VIRTQ_DESC_F_NEXT) |
| 354 | { |
| 355 | //Next one is chained |
| 356 | desc_index = desc_table[desc_index].next; |
| 357 | buffer_map_addr = desc_table[desc_index].addr; |
| 358 | buffer_len = desc_table[desc_index].len; |
| 359 | } |
| 360 | else if (vui->virtio_net_hdr_sz == 12) //MRG is available |
| 361 | { |
| 362 | virtio_net_hdr_mrg_rxbuf_t *hdr = |
| 363 | &vum->cpus[thread_index].tx_headers[tx_headers_len - 1]; |
| 364 | |
| 365 | //Move from available to used buffer |
| 366 | rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].id = |
| 367 | desc_head; |
| 368 | rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].len = |
| 369 | desc_len; |
| 370 | vhost_user_log_dirty_ring (vui, rxvq, |
| 371 | ring[rxvq->last_used_idx & |
| 372 | rxvq->qsz_mask]); |
| 373 | |
| 374 | rxvq->last_avail_idx++; |
| 375 | rxvq->last_used_idx++; |
| 376 | hdr->num_buffers++; |
| 377 | desc_len = 0; |
| 378 | |
| 379 | if (PREDICT_FALSE |
| 380 | (rxvq->last_avail_idx == rxvq->avail->idx)) |
| 381 | { |
| 382 | //Dequeue queued descriptors for this packet |
| 383 | rxvq->last_used_idx -= hdr->num_buffers - 1; |
| 384 | rxvq->last_avail_idx -= hdr->num_buffers - 1; |
| 385 | error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; |
| 386 | goto done; |
| 387 | } |
| 388 | |
| 389 | desc_table = rxvq->desc; |
| 390 | desc_head = desc_index = |
| 391 | rxvq->avail->ring[rxvq->last_avail_idx & rxvq->qsz_mask]; |
| 392 | if (PREDICT_FALSE |
| 393 | (rxvq->desc[desc_head].flags & VIRTQ_DESC_F_INDIRECT)) |
| 394 | { |
| 395 | //It is seriously unlikely that a driver will put indirect descriptor |
| 396 | //after non-indirect descriptor. |
| 397 | if (PREDICT_FALSE |
| 398 | (rxvq->desc[desc_head].len < sizeof (vring_desc_t))) |
| 399 | { |
| 400 | error = VHOST_USER_TX_FUNC_ERROR_INDIRECT_OVERFLOW; |
| 401 | goto done; |
| 402 | } |
| 403 | if (PREDICT_FALSE |
| 404 | (!(desc_table = |
| 405 | map_guest_mem (vui, |
| 406 | rxvq->desc[desc_index].addr, |
| 407 | &map_hint)))) |
| 408 | { |
| 409 | error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; |
| 410 | goto done; |
| 411 | } |
| 412 | desc_index = 0; |
| 413 | } |
| 414 | buffer_map_addr = desc_table[desc_index].addr; |
| 415 | buffer_len = desc_table[desc_index].len; |
| 416 | } |
| 417 | else |
| 418 | { |
| 419 | error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOMRG; |
| 420 | goto done; |
| 421 | } |
| 422 | } |
| 423 | |
| 424 | { |
| 425 | vhost_copy_t *cpy = &vum->cpus[thread_index].copy[copy_len]; |
| 426 | copy_len++; |
| 427 | cpy->len = bytes_left; |
| 428 | cpy->len = (cpy->len > buffer_len) ? buffer_len : cpy->len; |
| 429 | cpy->dst = buffer_map_addr; |
| 430 | cpy->src = (uword) vlib_buffer_get_current (current_b0) + |
| 431 | current_b0->current_length - bytes_left; |
| 432 | |
| 433 | bytes_left -= cpy->len; |
| 434 | buffer_len -= cpy->len; |
| 435 | buffer_map_addr += cpy->len; |
| 436 | desc_len += cpy->len; |
| 437 | |
| 438 | CLIB_PREFETCH (&rxvq->desc, CLIB_CACHE_LINE_BYTES, LOAD); |
| 439 | } |
| 440 | |
| 441 | // Check if vlib buffer has more data. If not, get more or break. |
| 442 | if (PREDICT_TRUE (!bytes_left)) |
| 443 | { |
| 444 | if (PREDICT_FALSE |
| 445 | (current_b0->flags & VLIB_BUFFER_NEXT_PRESENT)) |
| 446 | { |
| 447 | current_b0 = vlib_get_buffer (vm, current_b0->next_buffer); |
| 448 | bytes_left = current_b0->current_length; |
| 449 | } |
| 450 | else |
| 451 | { |
| 452 | //End of packet |
| 453 | break; |
| 454 | } |
| 455 | } |
| 456 | } |
| 457 | |
| 458 | //Move from available to used ring |
| 459 | rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].id = desc_head; |
| 460 | rxvq->used->ring[rxvq->last_used_idx & rxvq->qsz_mask].len = desc_len; |
| 461 | vhost_user_log_dirty_ring (vui, rxvq, |
| 462 | ring[rxvq->last_used_idx & rxvq->qsz_mask]); |
| 463 | rxvq->last_avail_idx++; |
| 464 | rxvq->last_used_idx++; |
| 465 | |
| 466 | if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED)) |
| 467 | { |
| 468 | vum->cpus[thread_index].current_trace->hdr = |
| 469 | vum->cpus[thread_index].tx_headers[tx_headers_len - 1]; |
| 470 | } |
| 471 | |
| 472 | n_left--; //At the end for error counting when 'goto done' is invoked |
| 473 | |
| 474 | /* |
| 475 | * Do the copy periodically to prevent |
| 476 | * vum->cpus[thread_index].copy array overflow and corrupt memory |
| 477 | */ |
| 478 | if (PREDICT_FALSE (copy_len >= VHOST_USER_TX_COPY_THRESHOLD)) |
| 479 | { |
| 480 | if (PREDICT_FALSE |
| 481 | (vhost_user_tx_copy (vui, vum->cpus[thread_index].copy, |
| 482 | copy_len, &map_hint))) |
| 483 | { |
| 484 | vlib_error_count (vm, node->node_index, |
| 485 | VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); |
| 486 | } |
| 487 | copy_len = 0; |
| 488 | |
| 489 | /* give buffers back to driver */ |
| 490 | CLIB_MEMORY_BARRIER (); |
| 491 | rxvq->used->idx = rxvq->last_used_idx; |
| 492 | vhost_user_log_dirty_ring (vui, rxvq, idx); |
| 493 | } |
| 494 | buffers++; |
| 495 | } |
| 496 | |
| 497 | done: |
| 498 | //Do the memory copies |
| 499 | if (PREDICT_FALSE |
| 500 | (vhost_user_tx_copy (vui, vum->cpus[thread_index].copy, |
| 501 | copy_len, &map_hint))) |
| 502 | { |
| 503 | vlib_error_count (vm, node->node_index, |
| 504 | VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL, 1); |
| 505 | } |
| 506 | |
| 507 | CLIB_MEMORY_BARRIER (); |
| 508 | rxvq->used->idx = rxvq->last_used_idx; |
| 509 | vhost_user_log_dirty_ring (vui, rxvq, idx); |
| 510 | |
| 511 | /* |
| 512 | * When n_left is set, error is always set to something too. |
| 513 | * In case error is due to lack of remaining buffers, we go back up and |
| 514 | * retry. |
| 515 | * The idea is that it is better to waste some time on packets |
| 516 | * that have been processed already than dropping them and get |
| 517 | * more fresh packets with a good likelyhood that they will be dropped too. |
| 518 | * This technique also gives more time to VM driver to pick-up packets. |
| 519 | * In case the traffic flows from physical to virtual interfaces, this |
| 520 | * technique will end-up leveraging the physical NIC buffer in order to |
| 521 | * absorb the VM's CPU jitter. |
| 522 | */ |
| 523 | if (n_left && (error == VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF) && retry) |
| 524 | { |
| 525 | retry--; |
| 526 | goto retry; |
| 527 | } |
| 528 | |
| 529 | /* interrupt (call) handling */ |
| 530 | if ((rxvq->callfd_idx != ~0) && |
| 531 | !(rxvq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) |
| 532 | { |
| 533 | rxvq->n_since_last_int += frame->n_vectors - n_left; |
| 534 | |
| 535 | if (rxvq->n_since_last_int > vum->coalesce_frames) |
| 536 | vhost_user_send_call (vm, rxvq); |
| 537 | } |
| 538 | |
| 539 | vhost_user_vring_unlock (vui, qid); |
| 540 | |
| 541 | done3: |
| 542 | if (PREDICT_FALSE (n_left && error != VHOST_USER_TX_FUNC_ERROR_NONE)) |
| 543 | { |
| 544 | vlib_error_count (vm, node->node_index, error, n_left); |
| 545 | vlib_increment_simple_counter |
| 546 | (vnet_main.interface_main.sw_if_counters |
| 547 | + VNET_INTERFACE_COUNTER_DROP, |
| 548 | thread_index, vui->sw_if_index, n_left); |
| 549 | } |
| 550 | |
| 551 | vlib_buffer_free (vm, vlib_frame_args (frame), frame->n_vectors); |
| 552 | return frame->n_vectors; |
| 553 | } |
| 554 | |
| 555 | static __clib_unused clib_error_t * |
| 556 | vhost_user_interface_rx_mode_change (vnet_main_t * vnm, u32 hw_if_index, |
| 557 | u32 qid, vnet_hw_interface_rx_mode mode) |
| 558 | { |
| 559 | vlib_main_t *vm = vnm->vlib_main; |
| 560 | vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); |
| 561 | vhost_user_main_t *vum = &vhost_user_main; |
| 562 | vhost_user_intf_t *vui = |
| 563 | pool_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); |
| 564 | vhost_user_vring_t *txvq = &vui->vrings[VHOST_VRING_IDX_TX (qid)]; |
| 565 | |
| 566 | if ((mode == VNET_HW_INTERFACE_RX_MODE_INTERRUPT) || |
| 567 | (mode == VNET_HW_INTERFACE_RX_MODE_ADAPTIVE)) |
| 568 | { |
| 569 | if (txvq->kickfd_idx == ~0) |
| 570 | { |
| 571 | // We cannot support interrupt mode if the driver opts out |
| 572 | return clib_error_return (0, "Driver does not support interrupt"); |
| 573 | } |
| 574 | if (txvq->mode == VNET_HW_INTERFACE_RX_MODE_POLLING) |
| 575 | { |
| 576 | vum->ifq_count++; |
| 577 | // Start the timer if this is the first encounter on interrupt |
| 578 | // interface/queue |
| 579 | if ((vum->ifq_count == 1) && |
| 580 | (vum->coalesce_time > 0.0) && (vum->coalesce_frames > 0)) |
| 581 | vlib_process_signal_event (vm, |
| 582 | vhost_user_send_interrupt_node.index, |
| 583 | VHOST_USER_EVENT_START_TIMER, 0); |
| 584 | } |
| 585 | } |
| 586 | else if (mode == VNET_HW_INTERFACE_RX_MODE_POLLING) |
| 587 | { |
| 588 | if (((txvq->mode == VNET_HW_INTERFACE_RX_MODE_INTERRUPT) || |
| 589 | (txvq->mode == VNET_HW_INTERFACE_RX_MODE_ADAPTIVE)) && |
| 590 | vum->ifq_count) |
| 591 | { |
| 592 | vum->ifq_count--; |
| 593 | // Stop the timer if there is no more interrupt interface/queue |
| 594 | if ((vum->ifq_count == 0) && |
| 595 | (vum->coalesce_time > 0.0) && (vum->coalesce_frames > 0)) |
| 596 | vlib_process_signal_event (vm, |
| 597 | vhost_user_send_interrupt_node.index, |
| 598 | VHOST_USER_EVENT_STOP_TIMER, 0); |
| 599 | } |
| 600 | } |
| 601 | |
| 602 | txvq->mode = mode; |
| 603 | if (mode == VNET_HW_INTERFACE_RX_MODE_POLLING) |
| 604 | txvq->used->flags = VRING_USED_F_NO_NOTIFY; |
| 605 | else if ((mode == VNET_HW_INTERFACE_RX_MODE_ADAPTIVE) || |
| 606 | (mode == VNET_HW_INTERFACE_RX_MODE_INTERRUPT)) |
| 607 | txvq->used->flags = 0; |
| 608 | else |
| 609 | { |
| 610 | clib_warning ("BUG: unhandled mode %d changed for if %d queue %d", mode, |
| 611 | hw_if_index, qid); |
| 612 | return clib_error_return (0, "unsupported"); |
| 613 | } |
| 614 | |
| 615 | return 0; |
| 616 | } |
| 617 | |
| 618 | static __clib_unused clib_error_t * |
| 619 | vhost_user_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, |
| 620 | u32 flags) |
| 621 | { |
| 622 | vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); |
| 623 | vhost_user_main_t *vum = &vhost_user_main; |
| 624 | vhost_user_intf_t *vui = |
| 625 | pool_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); |
| 626 | u32 hw_flags = 0; |
| 627 | vui->admin_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; |
| 628 | hw_flags = vui->admin_up ? VNET_HW_INTERFACE_FLAG_LINK_UP : 0; |
| 629 | |
| 630 | vnet_hw_interface_set_flags (vnm, vui->hw_if_index, hw_flags); |
| 631 | |
| 632 | return /* no error */ 0; |
| 633 | } |
| 634 | |
| 635 | #ifndef CLIB_MARCH_VARIANT |
| 636 | /* *INDENT-OFF* */ |
| 637 | VNET_DEVICE_CLASS (vhost_user_device_class) = { |
| 638 | .name = "vhost-user", |
| 639 | .tx_function = vhost_user_tx, |
| 640 | .tx_function_n_errors = VHOST_USER_TX_FUNC_N_ERROR, |
| 641 | .tx_function_error_strings = vhost_user_tx_func_error_strings, |
| 642 | .format_device_name = format_vhost_user_interface_name, |
| 643 | .name_renumber = vhost_user_name_renumber, |
| 644 | .admin_up_down_function = vhost_user_interface_admin_up_down, |
| 645 | .rx_mode_change_function = vhost_user_interface_rx_mode_change, |
| 646 | .format_tx_trace = format_vhost_trace, |
| 647 | }; |
| 648 | |
| 649 | #if __x86_64__ |
| 650 | vlib_node_function_t __clib_weak vhost_user_tx_avx512; |
| 651 | vlib_node_function_t __clib_weak vhost_user_tx_avx2; |
| 652 | static void __clib_constructor |
| 653 | vhost_user_tx_multiarch_select (void) |
| 654 | { |
| 655 | if (vhost_user_tx_avx512 && clib_cpu_supports_avx512f ()) |
| 656 | vhost_user_device_class.tx_function = vhost_user_tx_avx512; |
| 657 | else if (vhost_user_tx_avx2 && clib_cpu_supports_avx2 ()) |
| 658 | vhost_user_device_class.tx_function = vhost_user_tx_avx2; |
| 659 | } |
| 660 | #endif |
| 661 | #endif |
| 662 | |
| 663 | /* *INDENT-ON* */ |
| 664 | |
| 665 | /* |
| 666 | * fd.io coding-style-patch-verification: ON |
| 667 | * |
| 668 | * Local Variables: |
| 669 | * eval: (c-set-style "gnu") |
| 670 | * End: |
| 671 | */ |