1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (c) 2009, Microsoft Corporation. 4 * 5 * Authors: 6 * Haiyang Zhang <haiyangz@microsoft.com> 7 * Hank Janssen <hjanssen@microsoft.com> 8 */ 9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 10 11 #include <linux/kernel.h> 12 #include <linux/sched.h> 13 #include <linux/wait.h> 14 #include <linux/mm.h> 15 #include <linux/delay.h> 16 #include <linux/io.h> 17 #include <linux/slab.h> 18 #include <linux/netdevice.h> 19 #include <linux/if_ether.h> 20 #include <linux/vmalloc.h> 21 #include <linux/rtnetlink.h> 22 #include <linux/prefetch.h> 23 24 #include <asm/sync_bitops.h> 25 #include <asm/mshyperv.h> 26 27 #include "hyperv_net.h" 28 #include "netvsc_trace.h" 29 30 /* 31 * Switch the data path from the synthetic interface to the VF 32 * interface. 33 */ 34 int netvsc_switch_datapath(struct net_device *ndev, bool vf) 35 { 36 struct net_device_context *net_device_ctx = netdev_priv(ndev); 37 struct hv_device *dev = net_device_ctx->device_ctx; 38 struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev); 39 struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt; 40 int ret, retry = 0; 41 42 /* Block sending traffic to VF if it's about to be gone */ 43 if (!vf) 44 net_device_ctx->data_path_is_vf = vf; 45 46 memset(init_pkt, 0, sizeof(struct nvsp_message)); 47 init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH; 48 if (vf) 49 init_pkt->msg.v4_msg.active_dp.active_datapath = 50 NVSP_DATAPATH_VF; 51 else 52 init_pkt->msg.v4_msg.active_dp.active_datapath = 53 NVSP_DATAPATH_SYNTHETIC; 54 55 again: 56 trace_nvsp_send(ndev, init_pkt); 57 58 ret = vmbus_sendpacket(dev->channel, init_pkt, 59 sizeof(struct nvsp_message), 60 (unsigned long)init_pkt, VM_PKT_DATA_INBAND, 61 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); 62 63 /* If failed to switch to/from VF, let data_path_is_vf stay false, 64 * so we use synthetic path to send data. 65 */ 66 if (ret) { 67 if (ret != -EAGAIN) { 68 netdev_err(ndev, 69 "Unable to send sw datapath msg, err: %d\n", 70 ret); 71 return ret; 72 } 73 74 if (retry++ < RETRY_MAX) { 75 usleep_range(RETRY_US_LO, RETRY_US_HI); 76 goto again; 77 } else { 78 netdev_err( 79 ndev, 80 "Retry failed to send sw datapath msg, err: %d\n", 81 ret); 82 return ret; 83 } 84 } 85 86 wait_for_completion(&nv_dev->channel_init_wait); 87 net_device_ctx->data_path_is_vf = vf; 88 89 return 0; 90 } 91 92 /* Worker to setup sub channels on initial setup 93 * Initial hotplug event occurs in softirq context 94 * and can't wait for channels. 95 */ 96 static void netvsc_subchan_work(struct work_struct *w) 97 { 98 struct netvsc_device *nvdev = 99 container_of(w, struct netvsc_device, subchan_work); 100 struct rndis_device *rdev; 101 int i, ret; 102 103 /* Avoid deadlock with device removal already under RTNL */ 104 if (!rtnl_trylock()) { 105 schedule_work(w); 106 return; 107 } 108 109 rdev = nvdev->extension; 110 if (rdev) { 111 ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL); 112 if (ret == 0) { 113 netif_device_attach(rdev->ndev); 114 } else { 115 /* fallback to only primary channel */ 116 for (i = 1; i < nvdev->num_chn; i++) 117 netif_napi_del(&nvdev->chan_table[i].napi); 118 119 nvdev->max_chn = 1; 120 nvdev->num_chn = 1; 121 } 122 } 123 124 rtnl_unlock(); 125 } 126 127 static struct netvsc_device *alloc_net_device(void) 128 { 129 struct netvsc_device *net_device; 130 131 net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL); 132 if (!net_device) 133 return NULL; 134 135 init_waitqueue_head(&net_device->wait_drain); 136 net_device->destroy = false; 137 net_device->tx_disable = true; 138 139 net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT; 140 net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT; 141 142 init_completion(&net_device->channel_init_wait); 143 init_waitqueue_head(&net_device->subchan_open); 144 INIT_WORK(&net_device->subchan_work, netvsc_subchan_work); 145 146 return net_device; 147 } 148 149 static void free_netvsc_device(struct rcu_head *head) 150 { 151 struct netvsc_device *nvdev 152 = container_of(head, struct netvsc_device, rcu); 153 int i; 154 155 kfree(nvdev->extension); 156 157 if (nvdev->recv_original_buf) 158 vfree(nvdev->recv_original_buf); 159 else 160 vfree(nvdev->recv_buf); 161 162 if (nvdev->send_original_buf) 163 vfree(nvdev->send_original_buf); 164 else 165 vfree(nvdev->send_buf); 166 167 bitmap_free(nvdev->send_section_map); 168 169 for (i = 0; i < VRSS_CHANNEL_MAX; i++) { 170 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq); 171 kfree(nvdev->chan_table[i].recv_buf); 172 vfree(nvdev->chan_table[i].mrc.slots); 173 } 174 175 kfree(nvdev); 176 } 177 178 static void free_netvsc_device_rcu(struct netvsc_device *nvdev) 179 { 180 call_rcu(&nvdev->rcu, free_netvsc_device); 181 } 182 183 static void netvsc_revoke_recv_buf(struct hv_device *device, 184 struct netvsc_device *net_device, 185 struct net_device *ndev) 186 { 187 struct nvsp_message *revoke_packet; 188 int ret; 189 190 /* 191 * If we got a section count, it means we received a 192 * SendReceiveBufferComplete msg (ie sent 193 * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need 194 * to send a revoke msg here 195 */ 196 if (net_device->recv_section_cnt) { 197 /* Send the revoke receive buffer */ 198 revoke_packet = &net_device->revoke_packet; 199 memset(revoke_packet, 0, sizeof(struct nvsp_message)); 200 201 revoke_packet->hdr.msg_type = 202 NVSP_MSG1_TYPE_REVOKE_RECV_BUF; 203 revoke_packet->msg.v1_msg. 204 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID; 205 206 trace_nvsp_send(ndev, revoke_packet); 207 208 ret = vmbus_sendpacket(device->channel, 209 revoke_packet, 210 sizeof(struct nvsp_message), 211 VMBUS_RQST_ID_NO_RESPONSE, 212 VM_PKT_DATA_INBAND, 0); 213 /* If the failure is because the channel is rescinded; 214 * ignore the failure since we cannot send on a rescinded 215 * channel. This would allow us to properly cleanup 216 * even when the channel is rescinded. 217 */ 218 if (device->channel->rescind) 219 ret = 0; 220 /* 221 * If we failed here, we might as well return and 222 * have a leak rather than continue and a bugchk 223 */ 224 if (ret != 0) { 225 netdev_err(ndev, "unable to send " 226 "revoke receive buffer to netvsp\n"); 227 return; 228 } 229 net_device->recv_section_cnt = 0; 230 } 231 } 232 233 static void netvsc_revoke_send_buf(struct hv_device *device, 234 struct netvsc_device *net_device, 235 struct net_device *ndev) 236 { 237 struct nvsp_message *revoke_packet; 238 int ret; 239 240 /* Deal with the send buffer we may have setup. 241 * If we got a send section size, it means we received a 242 * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent 243 * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need 244 * to send a revoke msg here 245 */ 246 if (net_device->send_section_cnt) { 247 /* Send the revoke receive buffer */ 248 revoke_packet = &net_device->revoke_packet; 249 memset(revoke_packet, 0, sizeof(struct nvsp_message)); 250 251 revoke_packet->hdr.msg_type = 252 NVSP_MSG1_TYPE_REVOKE_SEND_BUF; 253 revoke_packet->msg.v1_msg.revoke_send_buf.id = 254 NETVSC_SEND_BUFFER_ID; 255 256 trace_nvsp_send(ndev, revoke_packet); 257 258 ret = vmbus_sendpacket(device->channel, 259 revoke_packet, 260 sizeof(struct nvsp_message), 261 VMBUS_RQST_ID_NO_RESPONSE, 262 VM_PKT_DATA_INBAND, 0); 263 264 /* If the failure is because the channel is rescinded; 265 * ignore the failure since we cannot send on a rescinded 266 * channel. This would allow us to properly cleanup 267 * even when the channel is rescinded. 268 */ 269 if (device->channel->rescind) 270 ret = 0; 271 272 /* If we failed here, we might as well return and 273 * have a leak rather than continue and a bugchk 274 */ 275 if (ret != 0) { 276 netdev_err(ndev, "unable to send " 277 "revoke send buffer to netvsp\n"); 278 return; 279 } 280 net_device->send_section_cnt = 0; 281 } 282 } 283 284 static void netvsc_teardown_recv_gpadl(struct hv_device *device, 285 struct netvsc_device *net_device, 286 struct net_device *ndev) 287 { 288 int ret; 289 290 if (net_device->recv_buf_gpadl_handle.gpadl_handle) { 291 ret = vmbus_teardown_gpadl(device->channel, 292 &net_device->recv_buf_gpadl_handle); 293 294 /* If we failed here, we might as well return and have a leak 295 * rather than continue and a bugchk 296 */ 297 if (ret != 0) { 298 netdev_err(ndev, 299 "unable to teardown receive buffer's gpadl\n"); 300 return; 301 } 302 } 303 } 304 305 static void netvsc_teardown_send_gpadl(struct hv_device *device, 306 struct netvsc_device *net_device, 307 struct net_device *ndev) 308 { 309 int ret; 310 311 if (net_device->send_buf_gpadl_handle.gpadl_handle) { 312 ret = vmbus_teardown_gpadl(device->channel, 313 &net_device->send_buf_gpadl_handle); 314 315 /* If we failed here, we might as well return and have a leak 316 * rather than continue and a bugchk 317 */ 318 if (ret != 0) { 319 netdev_err(ndev, 320 "unable to teardown send buffer's gpadl\n"); 321 return; 322 } 323 } 324 } 325 326 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx) 327 { 328 struct netvsc_channel *nvchan = &net_device->chan_table[q_idx]; 329 int node = cpu_to_node(nvchan->channel->target_cpu); 330 size_t size; 331 332 size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data); 333 nvchan->mrc.slots = vzalloc_node(size, node); 334 if (!nvchan->mrc.slots) 335 nvchan->mrc.slots = vzalloc(size); 336 337 return nvchan->mrc.slots ? 0 : -ENOMEM; 338 } 339 340 static int netvsc_init_buf(struct hv_device *device, 341 struct netvsc_device *net_device, 342 const struct netvsc_device_info *device_info) 343 { 344 struct nvsp_1_message_send_receive_buffer_complete *resp; 345 struct net_device *ndev = hv_get_drvdata(device); 346 struct nvsp_message *init_packet; 347 unsigned int buf_size; 348 int i, ret = 0; 349 void *vaddr; 350 351 /* Get receive buffer area. */ 352 buf_size = device_info->recv_sections * device_info->recv_section_size; 353 buf_size = roundup(buf_size, PAGE_SIZE); 354 355 /* Legacy hosts only allow smaller receive buffer */ 356 if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2) 357 buf_size = min_t(unsigned int, buf_size, 358 NETVSC_RECEIVE_BUFFER_SIZE_LEGACY); 359 360 net_device->recv_buf = vzalloc(buf_size); 361 if (!net_device->recv_buf) { 362 netdev_err(ndev, 363 "unable to allocate receive buffer of size %u\n", 364 buf_size); 365 ret = -ENOMEM; 366 goto cleanup; 367 } 368 369 net_device->recv_buf_size = buf_size; 370 371 /* 372 * Establish the gpadl handle for this buffer on this 373 * channel. Note: This call uses the vmbus connection rather 374 * than the channel to establish the gpadl handle. 375 */ 376 ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf, 377 buf_size, 378 &net_device->recv_buf_gpadl_handle); 379 if (ret != 0) { 380 netdev_err(ndev, 381 "unable to establish receive buffer's gpadl\n"); 382 goto cleanup; 383 } 384 385 if (hv_isolation_type_snp()) { 386 vaddr = hv_map_memory(net_device->recv_buf, buf_size); 387 if (!vaddr) { 388 ret = -ENOMEM; 389 goto cleanup; 390 } 391 392 net_device->recv_original_buf = net_device->recv_buf; 393 net_device->recv_buf = vaddr; 394 } 395 396 /* Notify the NetVsp of the gpadl handle */ 397 init_packet = &net_device->channel_init_pkt; 398 memset(init_packet, 0, sizeof(struct nvsp_message)); 399 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF; 400 init_packet->msg.v1_msg.send_recv_buf. 401 gpadl_handle = net_device->recv_buf_gpadl_handle.gpadl_handle; 402 init_packet->msg.v1_msg. 403 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID; 404 405 trace_nvsp_send(ndev, init_packet); 406 407 /* Send the gpadl notification request */ 408 ret = vmbus_sendpacket(device->channel, init_packet, 409 sizeof(struct nvsp_message), 410 (unsigned long)init_packet, 411 VM_PKT_DATA_INBAND, 412 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); 413 if (ret != 0) { 414 netdev_err(ndev, 415 "unable to send receive buffer's gpadl to netvsp\n"); 416 goto cleanup; 417 } 418 419 wait_for_completion(&net_device->channel_init_wait); 420 421 /* Check the response */ 422 resp = &init_packet->msg.v1_msg.send_recv_buf_complete; 423 if (resp->status != NVSP_STAT_SUCCESS) { 424 netdev_err(ndev, 425 "Unable to complete receive buffer initialization with NetVsp - status %d\n", 426 resp->status); 427 ret = -EINVAL; 428 goto cleanup; 429 } 430 431 /* Parse the response */ 432 netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n", 433 resp->num_sections, resp->sections[0].sub_alloc_size, 434 resp->sections[0].num_sub_allocs); 435 436 /* There should only be one section for the entire receive buffer */ 437 if (resp->num_sections != 1 || resp->sections[0].offset != 0) { 438 ret = -EINVAL; 439 goto cleanup; 440 } 441 442 net_device->recv_section_size = resp->sections[0].sub_alloc_size; 443 net_device->recv_section_cnt = resp->sections[0].num_sub_allocs; 444 445 /* Ensure buffer will not overflow */ 446 if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size * 447 (u64)net_device->recv_section_cnt > (u64)buf_size) { 448 netdev_err(ndev, "invalid recv_section_size %u\n", 449 net_device->recv_section_size); 450 ret = -EINVAL; 451 goto cleanup; 452 } 453 454 for (i = 0; i < VRSS_CHANNEL_MAX; i++) { 455 struct netvsc_channel *nvchan = &net_device->chan_table[i]; 456 457 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL); 458 if (nvchan->recv_buf == NULL) { 459 ret = -ENOMEM; 460 goto cleanup; 461 } 462 } 463 464 /* Setup receive completion ring. 465 * Add 1 to the recv_section_cnt because at least one entry in a 466 * ring buffer has to be empty. 467 */ 468 net_device->recv_completion_cnt = net_device->recv_section_cnt + 1; 469 ret = netvsc_alloc_recv_comp_ring(net_device, 0); 470 if (ret) 471 goto cleanup; 472 473 /* Now setup the send buffer. */ 474 buf_size = device_info->send_sections * device_info->send_section_size; 475 buf_size = round_up(buf_size, PAGE_SIZE); 476 477 net_device->send_buf = vzalloc(buf_size); 478 if (!net_device->send_buf) { 479 netdev_err(ndev, "unable to allocate send buffer of size %u\n", 480 buf_size); 481 ret = -ENOMEM; 482 goto cleanup; 483 } 484 net_device->send_buf_size = buf_size; 485 486 /* Establish the gpadl handle for this buffer on this 487 * channel. Note: This call uses the vmbus connection rather 488 * than the channel to establish the gpadl handle. 489 */ 490 ret = vmbus_establish_gpadl(device->channel, net_device->send_buf, 491 buf_size, 492 &net_device->send_buf_gpadl_handle); 493 if (ret != 0) { 494 netdev_err(ndev, 495 "unable to establish send buffer's gpadl\n"); 496 goto cleanup; 497 } 498 499 if (hv_isolation_type_snp()) { 500 vaddr = hv_map_memory(net_device->send_buf, buf_size); 501 if (!vaddr) { 502 ret = -ENOMEM; 503 goto cleanup; 504 } 505 506 net_device->send_original_buf = net_device->send_buf; 507 net_device->send_buf = vaddr; 508 } 509 510 /* Notify the NetVsp of the gpadl handle */ 511 init_packet = &net_device->channel_init_pkt; 512 memset(init_packet, 0, sizeof(struct nvsp_message)); 513 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF; 514 init_packet->msg.v1_msg.send_send_buf.gpadl_handle = 515 net_device->send_buf_gpadl_handle.gpadl_handle; 516 init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID; 517 518 trace_nvsp_send(ndev, init_packet); 519 520 /* Send the gpadl notification request */ 521 ret = vmbus_sendpacket(device->channel, init_packet, 522 sizeof(struct nvsp_message), 523 (unsigned long)init_packet, 524 VM_PKT_DATA_INBAND, 525 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); 526 if (ret != 0) { 527 netdev_err(ndev, 528 "unable to send send buffer's gpadl to netvsp\n"); 529 goto cleanup; 530 } 531 532 wait_for_completion(&net_device->channel_init_wait); 533 534 /* Check the response */ 535 if (init_packet->msg.v1_msg. 536 send_send_buf_complete.status != NVSP_STAT_SUCCESS) { 537 netdev_err(ndev, "Unable to complete send buffer " 538 "initialization with NetVsp - status %d\n", 539 init_packet->msg.v1_msg. 540 send_send_buf_complete.status); 541 ret = -EINVAL; 542 goto cleanup; 543 } 544 545 /* Parse the response */ 546 net_device->send_section_size = init_packet->msg. 547 v1_msg.send_send_buf_complete.section_size; 548 if (net_device->send_section_size < NETVSC_MTU_MIN) { 549 netdev_err(ndev, "invalid send_section_size %u\n", 550 net_device->send_section_size); 551 ret = -EINVAL; 552 goto cleanup; 553 } 554 555 /* Section count is simply the size divided by the section size. */ 556 net_device->send_section_cnt = buf_size / net_device->send_section_size; 557 558 netdev_dbg(ndev, "Send section size: %d, Section count:%d\n", 559 net_device->send_section_size, net_device->send_section_cnt); 560 561 /* Setup state for managing the send buffer. */ 562 net_device->send_section_map = bitmap_zalloc(net_device->send_section_cnt, 563 GFP_KERNEL); 564 if (!net_device->send_section_map) { 565 ret = -ENOMEM; 566 goto cleanup; 567 } 568 569 goto exit; 570 571 cleanup: 572 netvsc_revoke_recv_buf(device, net_device, ndev); 573 netvsc_revoke_send_buf(device, net_device, ndev); 574 netvsc_teardown_recv_gpadl(device, net_device, ndev); 575 netvsc_teardown_send_gpadl(device, net_device, ndev); 576 577 exit: 578 return ret; 579 } 580 581 /* Negotiate NVSP protocol version */ 582 static int negotiate_nvsp_ver(struct hv_device *device, 583 struct netvsc_device *net_device, 584 struct nvsp_message *init_packet, 585 u32 nvsp_ver) 586 { 587 struct net_device *ndev = hv_get_drvdata(device); 588 int ret; 589 590 memset(init_packet, 0, sizeof(struct nvsp_message)); 591 init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT; 592 init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver; 593 init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver; 594 trace_nvsp_send(ndev, init_packet); 595 596 /* Send the init request */ 597 ret = vmbus_sendpacket(device->channel, init_packet, 598 sizeof(struct nvsp_message), 599 (unsigned long)init_packet, 600 VM_PKT_DATA_INBAND, 601 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); 602 603 if (ret != 0) 604 return ret; 605 606 wait_for_completion(&net_device->channel_init_wait); 607 608 if (init_packet->msg.init_msg.init_complete.status != 609 NVSP_STAT_SUCCESS) 610 return -EINVAL; 611 612 if (nvsp_ver == NVSP_PROTOCOL_VERSION_1) 613 return 0; 614 615 /* NVSPv2 or later: Send NDIS config */ 616 memset(init_packet, 0, sizeof(struct nvsp_message)); 617 init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG; 618 init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN; 619 init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1; 620 621 if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) { 622 if (hv_is_isolation_supported()) 623 netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n"); 624 else 625 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1; 626 627 /* Teaming bit is needed to receive link speed updates */ 628 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1; 629 } 630 631 if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61) 632 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1; 633 634 trace_nvsp_send(ndev, init_packet); 635 636 ret = vmbus_sendpacket(device->channel, init_packet, 637 sizeof(struct nvsp_message), 638 VMBUS_RQST_ID_NO_RESPONSE, 639 VM_PKT_DATA_INBAND, 0); 640 641 return ret; 642 } 643 644 static int netvsc_connect_vsp(struct hv_device *device, 645 struct netvsc_device *net_device, 646 const struct netvsc_device_info *device_info) 647 { 648 struct net_device *ndev = hv_get_drvdata(device); 649 static const u32 ver_list[] = { 650 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2, 651 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5, 652 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61 653 }; 654 struct nvsp_message *init_packet; 655 int ndis_version, i, ret; 656 657 init_packet = &net_device->channel_init_pkt; 658 659 /* Negotiate the latest NVSP protocol supported */ 660 for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--) 661 if (negotiate_nvsp_ver(device, net_device, init_packet, 662 ver_list[i]) == 0) { 663 net_device->nvsp_version = ver_list[i]; 664 break; 665 } 666 667 if (i < 0) { 668 ret = -EPROTO; 669 goto cleanup; 670 } 671 672 if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) { 673 netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n", 674 net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61); 675 ret = -EPROTO; 676 goto cleanup; 677 } 678 679 pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version); 680 681 /* Send the ndis version */ 682 memset(init_packet, 0, sizeof(struct nvsp_message)); 683 684 if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4) 685 ndis_version = 0x00060001; 686 else 687 ndis_version = 0x0006001e; 688 689 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER; 690 init_packet->msg.v1_msg. 691 send_ndis_ver.ndis_major_ver = 692 (ndis_version & 0xFFFF0000) >> 16; 693 init_packet->msg.v1_msg. 694 send_ndis_ver.ndis_minor_ver = 695 ndis_version & 0xFFFF; 696 697 trace_nvsp_send(ndev, init_packet); 698 699 /* Send the init request */ 700 ret = vmbus_sendpacket(device->channel, init_packet, 701 sizeof(struct nvsp_message), 702 VMBUS_RQST_ID_NO_RESPONSE, 703 VM_PKT_DATA_INBAND, 0); 704 if (ret != 0) 705 goto cleanup; 706 707 708 ret = netvsc_init_buf(device, net_device, device_info); 709 710 cleanup: 711 return ret; 712 } 713 714 /* 715 * netvsc_device_remove - Callback when the root bus device is removed 716 */ 717 void netvsc_device_remove(struct hv_device *device) 718 { 719 struct net_device *ndev = hv_get_drvdata(device); 720 struct net_device_context *net_device_ctx = netdev_priv(ndev); 721 struct netvsc_device *net_device 722 = rtnl_dereference(net_device_ctx->nvdev); 723 int i; 724 725 /* 726 * Revoke receive buffer. If host is pre-Win2016 then tear down 727 * receive buffer GPADL. Do the same for send buffer. 728 */ 729 netvsc_revoke_recv_buf(device, net_device, ndev); 730 if (vmbus_proto_version < VERSION_WIN10) 731 netvsc_teardown_recv_gpadl(device, net_device, ndev); 732 733 netvsc_revoke_send_buf(device, net_device, ndev); 734 if (vmbus_proto_version < VERSION_WIN10) 735 netvsc_teardown_send_gpadl(device, net_device, ndev); 736 737 RCU_INIT_POINTER(net_device_ctx->nvdev, NULL); 738 739 /* Disable NAPI and disassociate its context from the device. */ 740 for (i = 0; i < net_device->num_chn; i++) { 741 /* See also vmbus_reset_channel_cb(). */ 742 napi_disable(&net_device->chan_table[i].napi); 743 netif_napi_del(&net_device->chan_table[i].napi); 744 } 745 746 /* 747 * At this point, no one should be accessing net_device 748 * except in here 749 */ 750 netdev_dbg(ndev, "net device safe to remove\n"); 751 752 /* Now, we can close the channel safely */ 753 vmbus_close(device->channel); 754 755 /* 756 * If host is Win2016 or higher then we do the GPADL tear down 757 * here after VMBus is closed. 758 */ 759 if (vmbus_proto_version >= VERSION_WIN10) { 760 netvsc_teardown_recv_gpadl(device, net_device, ndev); 761 netvsc_teardown_send_gpadl(device, net_device, ndev); 762 } 763 764 if (net_device->recv_original_buf) 765 hv_unmap_memory(net_device->recv_buf); 766 767 if (net_device->send_original_buf) 768 hv_unmap_memory(net_device->send_buf); 769 770 /* Release all resources */ 771 free_netvsc_device_rcu(net_device); 772 } 773 774 #define RING_AVAIL_PERCENT_HIWATER 20 775 #define RING_AVAIL_PERCENT_LOWATER 10 776 777 static inline void netvsc_free_send_slot(struct netvsc_device *net_device, 778 u32 index) 779 { 780 sync_change_bit(index, net_device->send_section_map); 781 } 782 783 static void netvsc_send_tx_complete(struct net_device *ndev, 784 struct netvsc_device *net_device, 785 struct vmbus_channel *channel, 786 const struct vmpacket_descriptor *desc, 787 int budget) 788 { 789 struct net_device_context *ndev_ctx = netdev_priv(ndev); 790 struct sk_buff *skb; 791 u16 q_idx = 0; 792 int queue_sends; 793 u64 cmd_rqst; 794 795 cmd_rqst = channel->request_addr_callback(channel, (u64)desc->trans_id); 796 if (cmd_rqst == VMBUS_RQST_ERROR) { 797 netdev_err(ndev, "Incorrect transaction id\n"); 798 return; 799 } 800 801 skb = (struct sk_buff *)(unsigned long)cmd_rqst; 802 803 /* Notify the layer above us */ 804 if (likely(skb)) { 805 struct hv_netvsc_packet *packet 806 = (struct hv_netvsc_packet *)skb->cb; 807 u32 send_index = packet->send_buf_index; 808 struct netvsc_stats *tx_stats; 809 810 if (send_index != NETVSC_INVALID_INDEX) 811 netvsc_free_send_slot(net_device, send_index); 812 q_idx = packet->q_idx; 813 814 tx_stats = &net_device->chan_table[q_idx].tx_stats; 815 816 u64_stats_update_begin(&tx_stats->syncp); 817 tx_stats->packets += packet->total_packets; 818 tx_stats->bytes += packet->total_bytes; 819 u64_stats_update_end(&tx_stats->syncp); 820 821 netvsc_dma_unmap(ndev_ctx->device_ctx, packet); 822 napi_consume_skb(skb, budget); 823 } 824 825 queue_sends = 826 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends); 827 828 if (unlikely(net_device->destroy)) { 829 if (queue_sends == 0) 830 wake_up(&net_device->wait_drain); 831 } else { 832 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx); 833 834 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable && 835 (hv_get_avail_to_write_percent(&channel->outbound) > 836 RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) { 837 netif_tx_wake_queue(txq); 838 ndev_ctx->eth_stats.wake_queue++; 839 } 840 } 841 } 842 843 static void netvsc_send_completion(struct net_device *ndev, 844 struct netvsc_device *net_device, 845 struct vmbus_channel *incoming_channel, 846 const struct vmpacket_descriptor *desc, 847 int budget) 848 { 849 const struct nvsp_message *nvsp_packet; 850 u32 msglen = hv_pkt_datalen(desc); 851 struct nvsp_message *pkt_rqst; 852 u64 cmd_rqst; 853 854 /* First check if this is a VMBUS completion without data payload */ 855 if (!msglen) { 856 cmd_rqst = incoming_channel->request_addr_callback(incoming_channel, 857 (u64)desc->trans_id); 858 if (cmd_rqst == VMBUS_RQST_ERROR) { 859 netdev_err(ndev, "Invalid transaction id\n"); 860 return; 861 } 862 863 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst; 864 switch (pkt_rqst->hdr.msg_type) { 865 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH: 866 complete(&net_device->channel_init_wait); 867 break; 868 869 default: 870 netdev_err(ndev, "Unexpected VMBUS completion!!\n"); 871 } 872 return; 873 } 874 875 /* Ensure packet is big enough to read header fields */ 876 if (msglen < sizeof(struct nvsp_message_header)) { 877 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen); 878 return; 879 } 880 881 nvsp_packet = hv_pkt_data(desc); 882 switch (nvsp_packet->hdr.msg_type) { 883 case NVSP_MSG_TYPE_INIT_COMPLETE: 884 if (msglen < sizeof(struct nvsp_message_header) + 885 sizeof(struct nvsp_message_init_complete)) { 886 netdev_err(ndev, "nvsp_msg length too small: %u\n", 887 msglen); 888 return; 889 } 890 fallthrough; 891 892 case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE: 893 if (msglen < sizeof(struct nvsp_message_header) + 894 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) { 895 netdev_err(ndev, "nvsp_msg1 length too small: %u\n", 896 msglen); 897 return; 898 } 899 fallthrough; 900 901 case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE: 902 if (msglen < sizeof(struct nvsp_message_header) + 903 sizeof(struct nvsp_1_message_send_send_buffer_complete)) { 904 netdev_err(ndev, "nvsp_msg1 length too small: %u\n", 905 msglen); 906 return; 907 } 908 fallthrough; 909 910 case NVSP_MSG5_TYPE_SUBCHANNEL: 911 if (msglen < sizeof(struct nvsp_message_header) + 912 sizeof(struct nvsp_5_subchannel_complete)) { 913 netdev_err(ndev, "nvsp_msg5 length too small: %u\n", 914 msglen); 915 return; 916 } 917 /* Copy the response back */ 918 memcpy(&net_device->channel_init_pkt, nvsp_packet, 919 sizeof(struct nvsp_message)); 920 complete(&net_device->channel_init_wait); 921 break; 922 923 case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE: 924 netvsc_send_tx_complete(ndev, net_device, incoming_channel, 925 desc, budget); 926 break; 927 928 default: 929 netdev_err(ndev, 930 "Unknown send completion type %d received!!\n", 931 nvsp_packet->hdr.msg_type); 932 } 933 } 934 935 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device) 936 { 937 unsigned long *map_addr = net_device->send_section_map; 938 unsigned int i; 939 940 for_each_clear_bit(i, map_addr, net_device->send_section_cnt) { 941 if (sync_test_and_set_bit(i, map_addr) == 0) 942 return i; 943 } 944 945 return NETVSC_INVALID_INDEX; 946 } 947 948 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device, 949 unsigned int section_index, 950 u32 pend_size, 951 struct hv_netvsc_packet *packet, 952 struct rndis_message *rndis_msg, 953 struct hv_page_buffer *pb, 954 bool xmit_more) 955 { 956 char *start = net_device->send_buf; 957 char *dest = start + (section_index * net_device->send_section_size) 958 + pend_size; 959 int i; 960 u32 padding = 0; 961 u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt : 962 packet->page_buf_cnt; 963 u32 remain; 964 965 /* Add padding */ 966 remain = packet->total_data_buflen & (net_device->pkt_align - 1); 967 if (xmit_more && remain) { 968 padding = net_device->pkt_align - remain; 969 rndis_msg->msg_len += padding; 970 packet->total_data_buflen += padding; 971 } 972 973 for (i = 0; i < page_count; i++) { 974 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT); 975 u32 offset = pb[i].offset; 976 u32 len = pb[i].len; 977 978 memcpy(dest, (src + offset), len); 979 dest += len; 980 } 981 982 if (padding) 983 memset(dest, 0, padding); 984 } 985 986 void netvsc_dma_unmap(struct hv_device *hv_dev, 987 struct hv_netvsc_packet *packet) 988 { 989 u32 page_count = packet->cp_partial ? 990 packet->page_buf_cnt - packet->rmsg_pgcnt : 991 packet->page_buf_cnt; 992 int i; 993 994 if (!hv_is_isolation_supported()) 995 return; 996 997 if (!packet->dma_range) 998 return; 999 1000 for (i = 0; i < page_count; i++) 1001 dma_unmap_single(&hv_dev->device, packet->dma_range[i].dma, 1002 packet->dma_range[i].mapping_size, 1003 DMA_TO_DEVICE); 1004 1005 kfree(packet->dma_range); 1006 } 1007 1008 /* netvsc_dma_map - Map swiotlb bounce buffer with data page of 1009 * packet sent by vmbus_sendpacket_pagebuffer() in the Isolation 1010 * VM. 1011 * 1012 * In isolation VM, netvsc send buffer has been marked visible to 1013 * host and so the data copied to send buffer doesn't need to use 1014 * bounce buffer. The data pages handled by vmbus_sendpacket_pagebuffer() 1015 * may not be copied to send buffer and so these pages need to be 1016 * mapped with swiotlb bounce buffer. netvsc_dma_map() is to do 1017 * that. The pfns in the struct hv_page_buffer need to be converted 1018 * to bounce buffer's pfn. The loop here is necessary because the 1019 * entries in the page buffer array are not necessarily full 1020 * pages of data. Each entry in the array has a separate offset and 1021 * len that may be non-zero, even for entries in the middle of the 1022 * array. And the entries are not physically contiguous. So each 1023 * entry must be individually mapped rather than as a contiguous unit. 1024 * So not use dma_map_sg() here. 1025 */ 1026 static int netvsc_dma_map(struct hv_device *hv_dev, 1027 struct hv_netvsc_packet *packet, 1028 struct hv_page_buffer *pb) 1029 { 1030 u32 page_count = packet->cp_partial ? 1031 packet->page_buf_cnt - packet->rmsg_pgcnt : 1032 packet->page_buf_cnt; 1033 dma_addr_t dma; 1034 int i; 1035 1036 if (!hv_is_isolation_supported()) 1037 return 0; 1038 1039 packet->dma_range = kcalloc(page_count, 1040 sizeof(*packet->dma_range), 1041 GFP_KERNEL); 1042 if (!packet->dma_range) 1043 return -ENOMEM; 1044 1045 for (i = 0; i < page_count; i++) { 1046 char *src = phys_to_virt((pb[i].pfn << HV_HYP_PAGE_SHIFT) 1047 + pb[i].offset); 1048 u32 len = pb[i].len; 1049 1050 dma = dma_map_single(&hv_dev->device, src, len, 1051 DMA_TO_DEVICE); 1052 if (dma_mapping_error(&hv_dev->device, dma)) { 1053 kfree(packet->dma_range); 1054 return -ENOMEM; 1055 } 1056 1057 /* pb[].offset and pb[].len are not changed during dma mapping 1058 * and so not reassign. 1059 */ 1060 packet->dma_range[i].dma = dma; 1061 packet->dma_range[i].mapping_size = len; 1062 pb[i].pfn = dma >> HV_HYP_PAGE_SHIFT; 1063 } 1064 1065 return 0; 1066 } 1067 1068 static inline int netvsc_send_pkt( 1069 struct hv_device *device, 1070 struct hv_netvsc_packet *packet, 1071 struct netvsc_device *net_device, 1072 struct hv_page_buffer *pb, 1073 struct sk_buff *skb) 1074 { 1075 struct nvsp_message nvmsg; 1076 struct nvsp_1_message_send_rndis_packet *rpkt = 1077 &nvmsg.msg.v1_msg.send_rndis_pkt; 1078 struct netvsc_channel * const nvchan = 1079 &net_device->chan_table[packet->q_idx]; 1080 struct vmbus_channel *out_channel = nvchan->channel; 1081 struct net_device *ndev = hv_get_drvdata(device); 1082 struct net_device_context *ndev_ctx = netdev_priv(ndev); 1083 struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx); 1084 u64 req_id; 1085 int ret; 1086 u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound); 1087 1088 memset(&nvmsg, 0, sizeof(struct nvsp_message)); 1089 nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT; 1090 if (skb) 1091 rpkt->channel_type = 0; /* 0 is RMC_DATA */ 1092 else 1093 rpkt->channel_type = 1; /* 1 is RMC_CONTROL */ 1094 1095 rpkt->send_buf_section_index = packet->send_buf_index; 1096 if (packet->send_buf_index == NETVSC_INVALID_INDEX) 1097 rpkt->send_buf_section_size = 0; 1098 else 1099 rpkt->send_buf_section_size = packet->total_data_buflen; 1100 1101 req_id = (ulong)skb; 1102 1103 if (out_channel->rescind) 1104 return -ENODEV; 1105 1106 trace_nvsp_send_pkt(ndev, out_channel, rpkt); 1107 1108 packet->dma_range = NULL; 1109 if (packet->page_buf_cnt) { 1110 if (packet->cp_partial) 1111 pb += packet->rmsg_pgcnt; 1112 1113 ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb); 1114 if (ret) { 1115 ret = -EAGAIN; 1116 goto exit; 1117 } 1118 1119 ret = vmbus_sendpacket_pagebuffer(out_channel, 1120 pb, packet->page_buf_cnt, 1121 &nvmsg, sizeof(nvmsg), 1122 req_id); 1123 1124 if (ret) 1125 netvsc_dma_unmap(ndev_ctx->device_ctx, packet); 1126 } else { 1127 ret = vmbus_sendpacket(out_channel, 1128 &nvmsg, sizeof(nvmsg), 1129 req_id, VM_PKT_DATA_INBAND, 1130 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); 1131 } 1132 1133 exit: 1134 if (ret == 0) { 1135 atomic_inc_return(&nvchan->queue_sends); 1136 1137 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) { 1138 netif_tx_stop_queue(txq); 1139 ndev_ctx->eth_stats.stop_queue++; 1140 } 1141 } else if (ret == -EAGAIN) { 1142 netif_tx_stop_queue(txq); 1143 ndev_ctx->eth_stats.stop_queue++; 1144 } else { 1145 netdev_err(ndev, 1146 "Unable to send packet pages %u len %u, ret %d\n", 1147 packet->page_buf_cnt, packet->total_data_buflen, 1148 ret); 1149 } 1150 1151 if (netif_tx_queue_stopped(txq) && 1152 atomic_read(&nvchan->queue_sends) < 1 && 1153 !net_device->tx_disable) { 1154 netif_tx_wake_queue(txq); 1155 ndev_ctx->eth_stats.wake_queue++; 1156 if (ret == -EAGAIN) 1157 ret = -ENOSPC; 1158 } 1159 1160 return ret; 1161 } 1162 1163 /* Move packet out of multi send data (msd), and clear msd */ 1164 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send, 1165 struct sk_buff **msd_skb, 1166 struct multi_send_data *msdp) 1167 { 1168 *msd_skb = msdp->skb; 1169 *msd_send = msdp->pkt; 1170 msdp->skb = NULL; 1171 msdp->pkt = NULL; 1172 msdp->count = 0; 1173 } 1174 1175 /* RCU already held by caller */ 1176 /* Batching/bouncing logic is designed to attempt to optimize 1177 * performance. 1178 * 1179 * For small, non-LSO packets we copy the packet to a send buffer 1180 * which is pre-registered with the Hyper-V side. This enables the 1181 * hypervisor to avoid remapping the aperture to access the packet 1182 * descriptor and data. 1183 * 1184 * If we already started using a buffer and the netdev is transmitting 1185 * a burst of packets, keep on copying into the buffer until it is 1186 * full or we are done collecting a burst. If there is an existing 1187 * buffer with space for the RNDIS descriptor but not the packet, copy 1188 * the RNDIS descriptor to the buffer, keeping the packet in place. 1189 * 1190 * If we do batching and send more than one packet using a single 1191 * NetVSC message, free the SKBs of the packets copied, except for the 1192 * last packet. This is done to streamline the handling of the case 1193 * where the last packet only had the RNDIS descriptor copied to the 1194 * send buffer, with the data pointers included in the NetVSC message. 1195 */ 1196 int netvsc_send(struct net_device *ndev, 1197 struct hv_netvsc_packet *packet, 1198 struct rndis_message *rndis_msg, 1199 struct hv_page_buffer *pb, 1200 struct sk_buff *skb, 1201 bool xdp_tx) 1202 { 1203 struct net_device_context *ndev_ctx = netdev_priv(ndev); 1204 struct netvsc_device *net_device 1205 = rcu_dereference_bh(ndev_ctx->nvdev); 1206 struct hv_device *device = ndev_ctx->device_ctx; 1207 int ret = 0; 1208 struct netvsc_channel *nvchan; 1209 u32 pktlen = packet->total_data_buflen, msd_len = 0; 1210 unsigned int section_index = NETVSC_INVALID_INDEX; 1211 struct multi_send_data *msdp; 1212 struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL; 1213 struct sk_buff *msd_skb = NULL; 1214 bool try_batch, xmit_more; 1215 1216 /* If device is rescinded, return error and packet will get dropped. */ 1217 if (unlikely(!net_device || net_device->destroy)) 1218 return -ENODEV; 1219 1220 nvchan = &net_device->chan_table[packet->q_idx]; 1221 packet->send_buf_index = NETVSC_INVALID_INDEX; 1222 packet->cp_partial = false; 1223 1224 /* Send a control message or XDP packet directly without accessing 1225 * msd (Multi-Send Data) field which may be changed during data packet 1226 * processing. 1227 */ 1228 if (!skb || xdp_tx) 1229 return netvsc_send_pkt(device, packet, net_device, pb, skb); 1230 1231 /* batch packets in send buffer if possible */ 1232 msdp = &nvchan->msd; 1233 if (msdp->pkt) 1234 msd_len = msdp->pkt->total_data_buflen; 1235 1236 try_batch = msd_len > 0 && msdp->count < net_device->max_pkt; 1237 if (try_batch && msd_len + pktlen + net_device->pkt_align < 1238 net_device->send_section_size) { 1239 section_index = msdp->pkt->send_buf_index; 1240 1241 } else if (try_batch && msd_len + packet->rmsg_size < 1242 net_device->send_section_size) { 1243 section_index = msdp->pkt->send_buf_index; 1244 packet->cp_partial = true; 1245 1246 } else if (pktlen + net_device->pkt_align < 1247 net_device->send_section_size) { 1248 section_index = netvsc_get_next_send_section(net_device); 1249 if (unlikely(section_index == NETVSC_INVALID_INDEX)) { 1250 ++ndev_ctx->eth_stats.tx_send_full; 1251 } else { 1252 move_pkt_msd(&msd_send, &msd_skb, msdp); 1253 msd_len = 0; 1254 } 1255 } 1256 1257 /* Keep aggregating only if stack says more data is coming 1258 * and not doing mixed modes send and not flow blocked 1259 */ 1260 xmit_more = netdev_xmit_more() && 1261 !packet->cp_partial && 1262 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx)); 1263 1264 if (section_index != NETVSC_INVALID_INDEX) { 1265 netvsc_copy_to_send_buf(net_device, 1266 section_index, msd_len, 1267 packet, rndis_msg, pb, xmit_more); 1268 1269 packet->send_buf_index = section_index; 1270 1271 if (packet->cp_partial) { 1272 packet->page_buf_cnt -= packet->rmsg_pgcnt; 1273 packet->total_data_buflen = msd_len + packet->rmsg_size; 1274 } else { 1275 packet->page_buf_cnt = 0; 1276 packet->total_data_buflen += msd_len; 1277 } 1278 1279 if (msdp->pkt) { 1280 packet->total_packets += msdp->pkt->total_packets; 1281 packet->total_bytes += msdp->pkt->total_bytes; 1282 } 1283 1284 if (msdp->skb) 1285 dev_consume_skb_any(msdp->skb); 1286 1287 if (xmit_more) { 1288 msdp->skb = skb; 1289 msdp->pkt = packet; 1290 msdp->count++; 1291 } else { 1292 cur_send = packet; 1293 msdp->skb = NULL; 1294 msdp->pkt = NULL; 1295 msdp->count = 0; 1296 } 1297 } else { 1298 move_pkt_msd(&msd_send, &msd_skb, msdp); 1299 cur_send = packet; 1300 } 1301 1302 if (msd_send) { 1303 int m_ret = netvsc_send_pkt(device, msd_send, net_device, 1304 NULL, msd_skb); 1305 1306 if (m_ret != 0) { 1307 netvsc_free_send_slot(net_device, 1308 msd_send->send_buf_index); 1309 dev_kfree_skb_any(msd_skb); 1310 } 1311 } 1312 1313 if (cur_send) 1314 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb); 1315 1316 if (ret != 0 && section_index != NETVSC_INVALID_INDEX) 1317 netvsc_free_send_slot(net_device, section_index); 1318 1319 return ret; 1320 } 1321 1322 /* Send pending recv completions */ 1323 static int send_recv_completions(struct net_device *ndev, 1324 struct netvsc_device *nvdev, 1325 struct netvsc_channel *nvchan) 1326 { 1327 struct multi_recv_comp *mrc = &nvchan->mrc; 1328 struct recv_comp_msg { 1329 struct nvsp_message_header hdr; 1330 u32 status; 1331 } __packed; 1332 struct recv_comp_msg msg = { 1333 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE, 1334 }; 1335 int ret; 1336 1337 while (mrc->first != mrc->next) { 1338 const struct recv_comp_data *rcd 1339 = mrc->slots + mrc->first; 1340 1341 msg.status = rcd->status; 1342 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg), 1343 rcd->tid, VM_PKT_COMP, 0); 1344 if (unlikely(ret)) { 1345 struct net_device_context *ndev_ctx = netdev_priv(ndev); 1346 1347 ++ndev_ctx->eth_stats.rx_comp_busy; 1348 return ret; 1349 } 1350 1351 if (++mrc->first == nvdev->recv_completion_cnt) 1352 mrc->first = 0; 1353 } 1354 1355 /* receive completion ring has been emptied */ 1356 if (unlikely(nvdev->destroy)) 1357 wake_up(&nvdev->wait_drain); 1358 1359 return 0; 1360 } 1361 1362 /* Count how many receive completions are outstanding */ 1363 static void recv_comp_slot_avail(const struct netvsc_device *nvdev, 1364 const struct multi_recv_comp *mrc, 1365 u32 *filled, u32 *avail) 1366 { 1367 u32 count = nvdev->recv_completion_cnt; 1368 1369 if (mrc->next >= mrc->first) 1370 *filled = mrc->next - mrc->first; 1371 else 1372 *filled = (count - mrc->first) + mrc->next; 1373 1374 *avail = count - *filled - 1; 1375 } 1376 1377 /* Add receive complete to ring to send to host. */ 1378 static void enq_receive_complete(struct net_device *ndev, 1379 struct netvsc_device *nvdev, u16 q_idx, 1380 u64 tid, u32 status) 1381 { 1382 struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx]; 1383 struct multi_recv_comp *mrc = &nvchan->mrc; 1384 struct recv_comp_data *rcd; 1385 u32 filled, avail; 1386 1387 recv_comp_slot_avail(nvdev, mrc, &filled, &avail); 1388 1389 if (unlikely(filled > NAPI_POLL_WEIGHT)) { 1390 send_recv_completions(ndev, nvdev, nvchan); 1391 recv_comp_slot_avail(nvdev, mrc, &filled, &avail); 1392 } 1393 1394 if (unlikely(!avail)) { 1395 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n", 1396 q_idx, tid); 1397 return; 1398 } 1399 1400 rcd = mrc->slots + mrc->next; 1401 rcd->tid = tid; 1402 rcd->status = status; 1403 1404 if (++mrc->next == nvdev->recv_completion_cnt) 1405 mrc->next = 0; 1406 } 1407 1408 static int netvsc_receive(struct net_device *ndev, 1409 struct netvsc_device *net_device, 1410 struct netvsc_channel *nvchan, 1411 const struct vmpacket_descriptor *desc) 1412 { 1413 struct net_device_context *net_device_ctx = netdev_priv(ndev); 1414 struct vmbus_channel *channel = nvchan->channel; 1415 const struct vmtransfer_page_packet_header *vmxferpage_packet 1416 = container_of(desc, const struct vmtransfer_page_packet_header, d); 1417 const struct nvsp_message *nvsp = hv_pkt_data(desc); 1418 u32 msglen = hv_pkt_datalen(desc); 1419 u16 q_idx = channel->offermsg.offer.sub_channel_index; 1420 char *recv_buf = net_device->recv_buf; 1421 u32 status = NVSP_STAT_SUCCESS; 1422 int i; 1423 int count = 0; 1424 1425 /* Ensure packet is big enough to read header fields */ 1426 if (msglen < sizeof(struct nvsp_message_header)) { 1427 netif_err(net_device_ctx, rx_err, ndev, 1428 "invalid nvsp header, length too small: %u\n", 1429 msglen); 1430 return 0; 1431 } 1432 1433 /* Make sure this is a valid nvsp packet */ 1434 if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) { 1435 netif_err(net_device_ctx, rx_err, ndev, 1436 "Unknown nvsp packet type received %u\n", 1437 nvsp->hdr.msg_type); 1438 return 0; 1439 } 1440 1441 /* Validate xfer page pkt header */ 1442 if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) { 1443 netif_err(net_device_ctx, rx_err, ndev, 1444 "Invalid xfer page pkt, offset too small: %u\n", 1445 desc->offset8 << 3); 1446 return 0; 1447 } 1448 1449 if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) { 1450 netif_err(net_device_ctx, rx_err, ndev, 1451 "Invalid xfer page set id - expecting %x got %x\n", 1452 NETVSC_RECEIVE_BUFFER_ID, 1453 vmxferpage_packet->xfer_pageset_id); 1454 return 0; 1455 } 1456 1457 count = vmxferpage_packet->range_cnt; 1458 1459 /* Check count for a valid value */ 1460 if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) { 1461 netif_err(net_device_ctx, rx_err, ndev, 1462 "Range count is not valid: %d\n", 1463 count); 1464 return 0; 1465 } 1466 1467 /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */ 1468 for (i = 0; i < count; i++) { 1469 u32 offset = vmxferpage_packet->ranges[i].byte_offset; 1470 u32 buflen = vmxferpage_packet->ranges[i].byte_count; 1471 void *data; 1472 int ret; 1473 1474 if (unlikely(offset > net_device->recv_buf_size || 1475 buflen > net_device->recv_buf_size - offset)) { 1476 nvchan->rsc.cnt = 0; 1477 status = NVSP_STAT_FAIL; 1478 netif_err(net_device_ctx, rx_err, ndev, 1479 "Packet offset:%u + len:%u too big\n", 1480 offset, buflen); 1481 1482 continue; 1483 } 1484 1485 /* We're going to copy (sections of) the packet into nvchan->recv_buf; 1486 * make sure that nvchan->recv_buf is large enough to hold the packet. 1487 */ 1488 if (unlikely(buflen > net_device->recv_section_size)) { 1489 nvchan->rsc.cnt = 0; 1490 status = NVSP_STAT_FAIL; 1491 netif_err(net_device_ctx, rx_err, ndev, 1492 "Packet too big: buflen=%u recv_section_size=%u\n", 1493 buflen, net_device->recv_section_size); 1494 1495 continue; 1496 } 1497 1498 data = recv_buf + offset; 1499 1500 nvchan->rsc.is_last = (i == count - 1); 1501 1502 trace_rndis_recv(ndev, q_idx, data); 1503 1504 /* Pass it to the upper layer */ 1505 ret = rndis_filter_receive(ndev, net_device, 1506 nvchan, data, buflen); 1507 1508 if (unlikely(ret != NVSP_STAT_SUCCESS)) { 1509 /* Drop incomplete packet */ 1510 nvchan->rsc.cnt = 0; 1511 status = NVSP_STAT_FAIL; 1512 } 1513 } 1514 1515 enq_receive_complete(ndev, net_device, q_idx, 1516 vmxferpage_packet->d.trans_id, status); 1517 1518 return count; 1519 } 1520 1521 static void netvsc_send_table(struct net_device *ndev, 1522 struct netvsc_device *nvscdev, 1523 const struct nvsp_message *nvmsg, 1524 u32 msglen) 1525 { 1526 struct net_device_context *net_device_ctx = netdev_priv(ndev); 1527 u32 count, offset, *tab; 1528 int i; 1529 1530 /* Ensure packet is big enough to read send_table fields */ 1531 if (msglen < sizeof(struct nvsp_message_header) + 1532 sizeof(struct nvsp_5_send_indirect_table)) { 1533 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen); 1534 return; 1535 } 1536 1537 count = nvmsg->msg.v5_msg.send_table.count; 1538 offset = nvmsg->msg.v5_msg.send_table.offset; 1539 1540 if (count != VRSS_SEND_TAB_SIZE) { 1541 netdev_err(ndev, "Received wrong send-table size:%u\n", count); 1542 return; 1543 } 1544 1545 /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be 1546 * wrong due to a host bug. So fix the offset here. 1547 */ 1548 if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 && 1549 msglen >= sizeof(struct nvsp_message_header) + 1550 sizeof(union nvsp_6_message_uber) + count * sizeof(u32)) 1551 offset = sizeof(struct nvsp_message_header) + 1552 sizeof(union nvsp_6_message_uber); 1553 1554 /* Boundary check for all versions */ 1555 if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) { 1556 netdev_err(ndev, "Received send-table offset too big:%u\n", 1557 offset); 1558 return; 1559 } 1560 1561 tab = (void *)nvmsg + offset; 1562 1563 for (i = 0; i < count; i++) 1564 net_device_ctx->tx_table[i] = tab[i]; 1565 } 1566 1567 static void netvsc_send_vf(struct net_device *ndev, 1568 const struct nvsp_message *nvmsg, 1569 u32 msglen) 1570 { 1571 struct net_device_context *net_device_ctx = netdev_priv(ndev); 1572 1573 /* Ensure packet is big enough to read its fields */ 1574 if (msglen < sizeof(struct nvsp_message_header) + 1575 sizeof(struct nvsp_4_send_vf_association)) { 1576 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen); 1577 return; 1578 } 1579 1580 net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated; 1581 net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial; 1582 netdev_info(ndev, "VF slot %u %s\n", 1583 net_device_ctx->vf_serial, 1584 net_device_ctx->vf_alloc ? "added" : "removed"); 1585 } 1586 1587 static void netvsc_receive_inband(struct net_device *ndev, 1588 struct netvsc_device *nvscdev, 1589 const struct vmpacket_descriptor *desc) 1590 { 1591 const struct nvsp_message *nvmsg = hv_pkt_data(desc); 1592 u32 msglen = hv_pkt_datalen(desc); 1593 1594 /* Ensure packet is big enough to read header fields */ 1595 if (msglen < sizeof(struct nvsp_message_header)) { 1596 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen); 1597 return; 1598 } 1599 1600 switch (nvmsg->hdr.msg_type) { 1601 case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE: 1602 netvsc_send_table(ndev, nvscdev, nvmsg, msglen); 1603 break; 1604 1605 case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION: 1606 if (hv_is_isolation_supported()) 1607 netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n"); 1608 else 1609 netvsc_send_vf(ndev, nvmsg, msglen); 1610 break; 1611 } 1612 } 1613 1614 static int netvsc_process_raw_pkt(struct hv_device *device, 1615 struct netvsc_channel *nvchan, 1616 struct netvsc_device *net_device, 1617 struct net_device *ndev, 1618 const struct vmpacket_descriptor *desc, 1619 int budget) 1620 { 1621 struct vmbus_channel *channel = nvchan->channel; 1622 const struct nvsp_message *nvmsg = hv_pkt_data(desc); 1623 1624 trace_nvsp_recv(ndev, channel, nvmsg); 1625 1626 switch (desc->type) { 1627 case VM_PKT_COMP: 1628 netvsc_send_completion(ndev, net_device, channel, desc, budget); 1629 break; 1630 1631 case VM_PKT_DATA_USING_XFER_PAGES: 1632 return netvsc_receive(ndev, net_device, nvchan, desc); 1633 1634 case VM_PKT_DATA_INBAND: 1635 netvsc_receive_inband(ndev, net_device, desc); 1636 break; 1637 1638 default: 1639 netdev_err(ndev, "unhandled packet type %d, tid %llx\n", 1640 desc->type, desc->trans_id); 1641 break; 1642 } 1643 1644 return 0; 1645 } 1646 1647 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel) 1648 { 1649 struct vmbus_channel *primary = channel->primary_channel; 1650 1651 return primary ? primary->device_obj : channel->device_obj; 1652 } 1653 1654 /* Network processing softirq 1655 * Process data in incoming ring buffer from host 1656 * Stops when ring is empty or budget is met or exceeded. 1657 */ 1658 int netvsc_poll(struct napi_struct *napi, int budget) 1659 { 1660 struct netvsc_channel *nvchan 1661 = container_of(napi, struct netvsc_channel, napi); 1662 struct netvsc_device *net_device = nvchan->net_device; 1663 struct vmbus_channel *channel = nvchan->channel; 1664 struct hv_device *device = netvsc_channel_to_device(channel); 1665 struct net_device *ndev = hv_get_drvdata(device); 1666 int work_done = 0; 1667 int ret; 1668 1669 /* If starting a new interval */ 1670 if (!nvchan->desc) 1671 nvchan->desc = hv_pkt_iter_first(channel); 1672 1673 while (nvchan->desc && work_done < budget) { 1674 work_done += netvsc_process_raw_pkt(device, nvchan, net_device, 1675 ndev, nvchan->desc, budget); 1676 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc); 1677 } 1678 1679 /* Send any pending receive completions */ 1680 ret = send_recv_completions(ndev, net_device, nvchan); 1681 1682 /* If it did not exhaust NAPI budget this time 1683 * and not doing busy poll 1684 * then re-enable host interrupts 1685 * and reschedule if ring is not empty 1686 * or sending receive completion failed. 1687 */ 1688 if (work_done < budget && 1689 napi_complete_done(napi, work_done) && 1690 (ret || hv_end_read(&channel->inbound)) && 1691 napi_schedule_prep(napi)) { 1692 hv_begin_read(&channel->inbound); 1693 __napi_schedule(napi); 1694 } 1695 1696 /* Driver may overshoot since multiple packets per descriptor */ 1697 return min(work_done, budget); 1698 } 1699 1700 /* Call back when data is available in host ring buffer. 1701 * Processing is deferred until network softirq (NAPI) 1702 */ 1703 void netvsc_channel_cb(void *context) 1704 { 1705 struct netvsc_channel *nvchan = context; 1706 struct vmbus_channel *channel = nvchan->channel; 1707 struct hv_ring_buffer_info *rbi = &channel->inbound; 1708 1709 /* preload first vmpacket descriptor */ 1710 prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index); 1711 1712 if (napi_schedule_prep(&nvchan->napi)) { 1713 /* disable interrupts from host */ 1714 hv_begin_read(rbi); 1715 1716 __napi_schedule_irqoff(&nvchan->napi); 1717 } 1718 } 1719 1720 /* 1721 * netvsc_device_add - Callback when the device belonging to this 1722 * driver is added 1723 */ 1724 struct netvsc_device *netvsc_device_add(struct hv_device *device, 1725 const struct netvsc_device_info *device_info) 1726 { 1727 int i, ret = 0; 1728 struct netvsc_device *net_device; 1729 struct net_device *ndev = hv_get_drvdata(device); 1730 struct net_device_context *net_device_ctx = netdev_priv(ndev); 1731 1732 net_device = alloc_net_device(); 1733 if (!net_device) 1734 return ERR_PTR(-ENOMEM); 1735 1736 for (i = 0; i < VRSS_SEND_TAB_SIZE; i++) 1737 net_device_ctx->tx_table[i] = 0; 1738 1739 /* Because the device uses NAPI, all the interrupt batching and 1740 * control is done via Net softirq, not the channel handling 1741 */ 1742 set_channel_read_mode(device->channel, HV_CALL_ISR); 1743 1744 /* If we're reopening the device we may have multiple queues, fill the 1745 * chn_table with the default channel to use it before subchannels are 1746 * opened. 1747 * Initialize the channel state before we open; 1748 * we can be interrupted as soon as we open the channel. 1749 */ 1750 1751 for (i = 0; i < VRSS_CHANNEL_MAX; i++) { 1752 struct netvsc_channel *nvchan = &net_device->chan_table[i]; 1753 1754 nvchan->channel = device->channel; 1755 nvchan->net_device = net_device; 1756 u64_stats_init(&nvchan->tx_stats.syncp); 1757 u64_stats_init(&nvchan->rx_stats.syncp); 1758 1759 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0); 1760 1761 if (ret) { 1762 netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret); 1763 goto cleanup2; 1764 } 1765 1766 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq, 1767 MEM_TYPE_PAGE_SHARED, NULL); 1768 1769 if (ret) { 1770 netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret); 1771 goto cleanup2; 1772 } 1773 } 1774 1775 /* Enable NAPI handler before init callbacks */ 1776 netif_napi_add(ndev, &net_device->chan_table[0].napi, 1777 netvsc_poll, NAPI_POLL_WEIGHT); 1778 1779 /* Open the channel */ 1780 device->channel->next_request_id_callback = vmbus_next_request_id; 1781 device->channel->request_addr_callback = vmbus_request_addr; 1782 device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes); 1783 device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE; 1784 1785 ret = vmbus_open(device->channel, netvsc_ring_bytes, 1786 netvsc_ring_bytes, NULL, 0, 1787 netvsc_channel_cb, net_device->chan_table); 1788 1789 if (ret != 0) { 1790 netdev_err(ndev, "unable to open channel: %d\n", ret); 1791 goto cleanup; 1792 } 1793 1794 /* Channel is opened */ 1795 netdev_dbg(ndev, "hv_netvsc channel opened successfully\n"); 1796 1797 napi_enable(&net_device->chan_table[0].napi); 1798 1799 /* Connect with the NetVsp */ 1800 ret = netvsc_connect_vsp(device, net_device, device_info); 1801 if (ret != 0) { 1802 netdev_err(ndev, 1803 "unable to connect to NetVSP - %d\n", ret); 1804 goto close; 1805 } 1806 1807 /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is 1808 * populated. 1809 */ 1810 rcu_assign_pointer(net_device_ctx->nvdev, net_device); 1811 1812 return net_device; 1813 1814 close: 1815 RCU_INIT_POINTER(net_device_ctx->nvdev, NULL); 1816 napi_disable(&net_device->chan_table[0].napi); 1817 1818 /* Now, we can close the channel safely */ 1819 vmbus_close(device->channel); 1820 1821 cleanup: 1822 netif_napi_del(&net_device->chan_table[0].napi); 1823 1824 cleanup2: 1825 if (net_device->recv_original_buf) 1826 hv_unmap_memory(net_device->recv_buf); 1827 1828 if (net_device->send_original_buf) 1829 hv_unmap_memory(net_device->send_buf); 1830 1831 free_netvsc_device(&net_device->rcu); 1832 1833 return ERR_PTR(ret); 1834 } 1835