1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Copyright (C) 2017 - 2019 Cambridge Greys Limited 4 * Copyright (C) 2011 - 2014 Cisco Systems Inc 5 * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com) 6 * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and 7 * James Leu (jleu@mindspring.net). 8 * Copyright (C) 2001 by various other people who didn't put their name here. 9 */ 10 11#include <linux/version.h> 12#include <linux/memblock.h> 13#include <linux/etherdevice.h> 14#include <linux/ethtool.h> 15#include <linux/inetdevice.h> 16#include <linux/init.h> 17#include <linux/list.h> 18#include <linux/netdevice.h> 19#include <linux/platform_device.h> 20#include <linux/rtnetlink.h> 21#include <linux/skbuff.h> 22#include <linux/slab.h> 23#include <linux/interrupt.h> 24#include <linux/firmware.h> 25#include <linux/fs.h> 26#include <uapi/linux/filter.h> 27#include <init.h> 28#include <irq_kern.h> 29#include <irq_user.h> 30#include <net_kern.h> 31#include <os.h> 32#include "mconsole_kern.h" 33#include "vector_user.h" 34#include "vector_kern.h" 35 36/* 37 * Adapted from network devices with the following major changes: 38 * All transports are static - simplifies the code significantly 39 * Multiple FDs/IRQs per device 40 * Vector IO optionally used for read/write, falling back to legacy 41 * based on configuration and/or availability 42 * Configuration is no longer positional - L2TPv3 and GRE require up to 43 * 10 parameters, passing this as positional is not fit for purpose. 44 * Only socket transports are supported 45 */ 46 47 48#define DRIVER_NAME "uml-vector" 49struct vector_cmd_line_arg { 50 struct list_head list; 51 int unit; 52 char *arguments; 53}; 54 55struct vector_device { 56 struct list_head list; 57 struct net_device *dev; 58 struct platform_device pdev; 59 int unit; 60 int opened; 61}; 62 63static LIST_HEAD(vec_cmd_line); 64 65static DEFINE_SPINLOCK(vector_devices_lock); 66static LIST_HEAD(vector_devices); 67 68static int driver_registered; 69 70static void vector_eth_configure(int n, struct arglist *def); 71 72/* Argument accessors to set variables (and/or set default values) 73 * mtu, buffer sizing, default headroom, etc 74 */ 75 76#define DEFAULT_HEADROOM 2 77#define SAFETY_MARGIN 32 78#define DEFAULT_VECTOR_SIZE 64 79#define TX_SMALL_PACKET 128 80#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1) 81#define MAX_ITERATIONS 64 82 83static const struct { 84 const char string[ETH_GSTRING_LEN]; 85} ethtool_stats_keys[] = { 86 { "rx_queue_max" }, 87 { "rx_queue_running_average" }, 88 { "tx_queue_max" }, 89 { "tx_queue_running_average" }, 90 { "rx_encaps_errors" }, 91 { "tx_timeout_count" }, 92 { "tx_restart_queue" }, 93 { "tx_kicks" }, 94 { "tx_flow_control_xon" }, 95 { "tx_flow_control_xoff" }, 96 { "rx_csum_offload_good" }, 97 { "rx_csum_offload_errors"}, 98 { "sg_ok"}, 99 { "sg_linearized"}, 100}; 101 102#define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys) 103 104static void vector_reset_stats(struct vector_private *vp) 105{ 106 vp->estats.rx_queue_max = 0; 107 vp->estats.rx_queue_running_average = 0; 108 vp->estats.tx_queue_max = 0; 109 vp->estats.tx_queue_running_average = 0; 110 vp->estats.rx_encaps_errors = 0; 111 vp->estats.tx_timeout_count = 0; 112 vp->estats.tx_restart_queue = 0; 113 vp->estats.tx_kicks = 0; 114 vp->estats.tx_flow_control_xon = 0; 115 vp->estats.tx_flow_control_xoff = 0; 116 vp->estats.sg_ok = 0; 117 vp->estats.sg_linearized = 0; 118} 119 120static int get_mtu(struct arglist *def) 121{ 122 char *mtu = uml_vector_fetch_arg(def, "mtu"); 123 long result; 124 125 if (mtu != NULL) { 126 if (kstrtoul(mtu, 10, &result) == 0) 127 if ((result < (1 << 16) - 1) && (result >= 576)) 128 return result; 129 } 130 return ETH_MAX_PACKET; 131} 132 133static char *get_bpf_file(struct arglist *def) 134{ 135 return uml_vector_fetch_arg(def, "bpffile"); 136} 137 138static bool get_bpf_flash(struct arglist *def) 139{ 140 char *allow = uml_vector_fetch_arg(def, "bpfflash"); 141 long result; 142 143 if (allow != NULL) { 144 if (kstrtoul(allow, 10, &result) == 0) 145 return (allow > 0); 146 } 147 return false; 148} 149 150static int get_depth(struct arglist *def) 151{ 152 char *mtu = uml_vector_fetch_arg(def, "depth"); 153 long result; 154 155 if (mtu != NULL) { 156 if (kstrtoul(mtu, 10, &result) == 0) 157 return result; 158 } 159 return DEFAULT_VECTOR_SIZE; 160} 161 162static int get_headroom(struct arglist *def) 163{ 164 char *mtu = uml_vector_fetch_arg(def, "headroom"); 165 long result; 166 167 if (mtu != NULL) { 168 if (kstrtoul(mtu, 10, &result) == 0) 169 return result; 170 } 171 return DEFAULT_HEADROOM; 172} 173 174static int get_req_size(struct arglist *def) 175{ 176 char *gro = uml_vector_fetch_arg(def, "gro"); 177 long result; 178 179 if (gro != NULL) { 180 if (kstrtoul(gro, 10, &result) == 0) { 181 if (result > 0) 182 return 65536; 183 } 184 } 185 return get_mtu(def) + ETH_HEADER_OTHER + 186 get_headroom(def) + SAFETY_MARGIN; 187} 188 189 190static int get_transport_options(struct arglist *def) 191{ 192 char *transport = uml_vector_fetch_arg(def, "transport"); 193 char *vector = uml_vector_fetch_arg(def, "vec"); 194 195 int vec_rx = VECTOR_RX; 196 int vec_tx = VECTOR_TX; 197 long parsed; 198 int result = 0; 199 200 if (transport == NULL) 201 return -EINVAL; 202 203 if (vector != NULL) { 204 if (kstrtoul(vector, 10, &parsed) == 0) { 205 if (parsed == 0) { 206 vec_rx = 0; 207 vec_tx = 0; 208 } 209 } 210 } 211 212 if (get_bpf_flash(def)) 213 result = VECTOR_BPF_FLASH; 214 215 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0) 216 return result; 217 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0) 218 return (result | vec_rx | VECTOR_BPF); 219 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0) 220 return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS); 221 return (result | vec_rx | vec_tx); 222} 223 224 225/* A mini-buffer for packet drop read 226 * All of our supported transports are datagram oriented and we always 227 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller 228 * than the packet size it still counts as full packet read and will 229 * clean the incoming stream to keep sigio/epoll happy 230 */ 231 232#define DROP_BUFFER_SIZE 32 233 234static char *drop_buffer; 235 236/* Array backed queues optimized for bulk enqueue/dequeue and 237 * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios. 238 * For more details and full design rationale see 239 * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt 240 */ 241 242 243/* 244 * Advance the mmsg queue head by n = advance. Resets the queue to 245 * maximum enqueue/dequeue-at-once capacity if possible. Called by 246 * dequeuers. Caller must hold the head_lock! 247 */ 248 249static int vector_advancehead(struct vector_queue *qi, int advance) 250{ 251 int queue_depth; 252 253 qi->head = 254 (qi->head + advance) 255 % qi->max_depth; 256 257 258 spin_lock(&qi->tail_lock); 259 qi->queue_depth -= advance; 260 261 /* we are at 0, use this to 262 * reset head and tail so we can use max size vectors 263 */ 264 265 if (qi->queue_depth == 0) { 266 qi->head = 0; 267 qi->tail = 0; 268 } 269 queue_depth = qi->queue_depth; 270 spin_unlock(&qi->tail_lock); 271 return queue_depth; 272} 273 274/* Advance the queue tail by n = advance. 275 * This is called by enqueuers which should hold the 276 * head lock already 277 */ 278 279static int vector_advancetail(struct vector_queue *qi, int advance) 280{ 281 int queue_depth; 282 283 qi->tail = 284 (qi->tail + advance) 285 % qi->max_depth; 286 spin_lock(&qi->head_lock); 287 qi->queue_depth += advance; 288 queue_depth = qi->queue_depth; 289 spin_unlock(&qi->head_lock); 290 return queue_depth; 291} 292 293static int prep_msg(struct vector_private *vp, 294 struct sk_buff *skb, 295 struct iovec *iov) 296{ 297 int iov_index = 0; 298 int nr_frags, frag; 299 skb_frag_t *skb_frag; 300 301 nr_frags = skb_shinfo(skb)->nr_frags; 302 if (nr_frags > MAX_IOV_SIZE) { 303 if (skb_linearize(skb) != 0) 304 goto drop; 305 } 306 if (vp->header_size > 0) { 307 iov[iov_index].iov_len = vp->header_size; 308 vp->form_header(iov[iov_index].iov_base, skb, vp); 309 iov_index++; 310 } 311 iov[iov_index].iov_base = skb->data; 312 if (nr_frags > 0) { 313 iov[iov_index].iov_len = skb->len - skb->data_len; 314 vp->estats.sg_ok++; 315 } else 316 iov[iov_index].iov_len = skb->len; 317 iov_index++; 318 for (frag = 0; frag < nr_frags; frag++) { 319 skb_frag = &skb_shinfo(skb)->frags[frag]; 320 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag); 321 iov[iov_index].iov_len = skb_frag_size(skb_frag); 322 iov_index++; 323 } 324 return iov_index; 325drop: 326 return -1; 327} 328/* 329 * Generic vector enqueue with support for forming headers using transport 330 * specific callback. Allows GRE, L2TPv3, RAW and other transports 331 * to use a common enqueue procedure in vector mode 332 */ 333 334static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb) 335{ 336 struct vector_private *vp = netdev_priv(qi->dev); 337 int queue_depth; 338 int packet_len; 339 struct mmsghdr *mmsg_vector = qi->mmsg_vector; 340 int iov_count; 341 342 spin_lock(&qi->tail_lock); 343 spin_lock(&qi->head_lock); 344 queue_depth = qi->queue_depth; 345 spin_unlock(&qi->head_lock); 346 347 if (skb) 348 packet_len = skb->len; 349 350 if (queue_depth < qi->max_depth) { 351 352 *(qi->skbuff_vector + qi->tail) = skb; 353 mmsg_vector += qi->tail; 354 iov_count = prep_msg( 355 vp, 356 skb, 357 mmsg_vector->msg_hdr.msg_iov 358 ); 359 if (iov_count < 1) 360 goto drop; 361 mmsg_vector->msg_hdr.msg_iovlen = iov_count; 362 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr; 363 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size; 364 queue_depth = vector_advancetail(qi, 1); 365 } else 366 goto drop; 367 spin_unlock(&qi->tail_lock); 368 return queue_depth; 369drop: 370 qi->dev->stats.tx_dropped++; 371 if (skb != NULL) { 372 packet_len = skb->len; 373 dev_consume_skb_any(skb); 374 netdev_completed_queue(qi->dev, 1, packet_len); 375 } 376 spin_unlock(&qi->tail_lock); 377 return queue_depth; 378} 379 380static int consume_vector_skbs(struct vector_queue *qi, int count) 381{ 382 struct sk_buff *skb; 383 int skb_index; 384 int bytes_compl = 0; 385 386 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) { 387 skb = *(qi->skbuff_vector + skb_index); 388 /* mark as empty to ensure correct destruction if 389 * needed 390 */ 391 bytes_compl += skb->len; 392 *(qi->skbuff_vector + skb_index) = NULL; 393 dev_consume_skb_any(skb); 394 } 395 qi->dev->stats.tx_bytes += bytes_compl; 396 qi->dev->stats.tx_packets += count; 397 netdev_completed_queue(qi->dev, count, bytes_compl); 398 return vector_advancehead(qi, count); 399} 400 401/* 402 * Generic vector deque via sendmmsg with support for forming headers 403 * using transport specific callback. Allows GRE, L2TPv3, RAW and 404 * other transports to use a common dequeue procedure in vector mode 405 */ 406 407 408static int vector_send(struct vector_queue *qi) 409{ 410 struct vector_private *vp = netdev_priv(qi->dev); 411 struct mmsghdr *send_from; 412 int result = 0, send_len, queue_depth = qi->max_depth; 413 414 if (spin_trylock(&qi->head_lock)) { 415 if (spin_trylock(&qi->tail_lock)) { 416 /* update queue_depth to current value */ 417 queue_depth = qi->queue_depth; 418 spin_unlock(&qi->tail_lock); 419 while (queue_depth > 0) { 420 /* Calculate the start of the vector */ 421 send_len = queue_depth; 422 send_from = qi->mmsg_vector; 423 send_from += qi->head; 424 /* Adjust vector size if wraparound */ 425 if (send_len + qi->head > qi->max_depth) 426 send_len = qi->max_depth - qi->head; 427 /* Try to TX as many packets as possible */ 428 if (send_len > 0) { 429 result = uml_vector_sendmmsg( 430 vp->fds->tx_fd, 431 send_from, 432 send_len, 433 0 434 ); 435 vp->in_write_poll = 436 (result != send_len); 437 } 438 /* For some of the sendmmsg error scenarios 439 * we may end being unsure in the TX success 440 * for all packets. It is safer to declare 441 * them all TX-ed and blame the network. 442 */ 443 if (result < 0) { 444 if (net_ratelimit()) 445 netdev_err(vp->dev, "sendmmsg err=%i\n", 446 result); 447 vp->in_error = true; 448 result = send_len; 449 } 450 if (result > 0) { 451 queue_depth = 452 consume_vector_skbs(qi, result); 453 /* This is equivalent to an TX IRQ. 454 * Restart the upper layers to feed us 455 * more packets. 456 */ 457 if (result > vp->estats.tx_queue_max) 458 vp->estats.tx_queue_max = result; 459 vp->estats.tx_queue_running_average = 460 (vp->estats.tx_queue_running_average + result) >> 1; 461 } 462 netif_trans_update(qi->dev); 463 netif_wake_queue(qi->dev); 464 /* if TX is busy, break out of the send loop, 465 * poll write IRQ will reschedule xmit for us 466 */ 467 if (result != send_len) { 468 vp->estats.tx_restart_queue++; 469 break; 470 } 471 } 472 } 473 spin_unlock(&qi->head_lock); 474 } else { 475 tasklet_schedule(&vp->tx_poll); 476 } 477 return queue_depth; 478} 479 480/* Queue destructor. Deliberately stateless so we can use 481 * it in queue cleanup if initialization fails. 482 */ 483 484static void destroy_queue(struct vector_queue *qi) 485{ 486 int i; 487 struct iovec *iov; 488 struct vector_private *vp = netdev_priv(qi->dev); 489 struct mmsghdr *mmsg_vector; 490 491 if (qi == NULL) 492 return; 493 /* deallocate any skbuffs - we rely on any unused to be 494 * set to NULL. 495 */ 496 if (qi->skbuff_vector != NULL) { 497 for (i = 0; i < qi->max_depth; i++) { 498 if (*(qi->skbuff_vector + i) != NULL) 499 dev_kfree_skb_any(*(qi->skbuff_vector + i)); 500 } 501 kfree(qi->skbuff_vector); 502 } 503 /* deallocate matching IOV structures including header buffs */ 504 if (qi->mmsg_vector != NULL) { 505 mmsg_vector = qi->mmsg_vector; 506 for (i = 0; i < qi->max_depth; i++) { 507 iov = mmsg_vector->msg_hdr.msg_iov; 508 if (iov != NULL) { 509 if ((vp->header_size > 0) && 510 (iov->iov_base != NULL)) 511 kfree(iov->iov_base); 512 kfree(iov); 513 } 514 mmsg_vector++; 515 } 516 kfree(qi->mmsg_vector); 517 } 518 kfree(qi); 519} 520 521/* 522 * Queue constructor. Create a queue with a given side. 523 */ 524static struct vector_queue *create_queue( 525 struct vector_private *vp, 526 int max_size, 527 int header_size, 528 int num_extra_frags) 529{ 530 struct vector_queue *result; 531 int i; 532 struct iovec *iov; 533 struct mmsghdr *mmsg_vector; 534 535 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL); 536 if (result == NULL) 537 return NULL; 538 result->max_depth = max_size; 539 result->dev = vp->dev; 540 result->mmsg_vector = kmalloc( 541 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL); 542 if (result->mmsg_vector == NULL) 543 goto out_mmsg_fail; 544 result->skbuff_vector = kmalloc( 545 (sizeof(void *) * max_size), GFP_KERNEL); 546 if (result->skbuff_vector == NULL) 547 goto out_skb_fail; 548 549 /* further failures can be handled safely by destroy_queue*/ 550 551 mmsg_vector = result->mmsg_vector; 552 for (i = 0; i < max_size; i++) { 553 /* Clear all pointers - we use non-NULL as marking on 554 * what to free on destruction 555 */ 556 *(result->skbuff_vector + i) = NULL; 557 mmsg_vector->msg_hdr.msg_iov = NULL; 558 mmsg_vector++; 559 } 560 mmsg_vector = result->mmsg_vector; 561 result->max_iov_frags = num_extra_frags; 562 for (i = 0; i < max_size; i++) { 563 if (vp->header_size > 0) 564 iov = kmalloc_array(3 + num_extra_frags, 565 sizeof(struct iovec), 566 GFP_KERNEL 567 ); 568 else 569 iov = kmalloc_array(2 + num_extra_frags, 570 sizeof(struct iovec), 571 GFP_KERNEL 572 ); 573 if (iov == NULL) 574 goto out_fail; 575 mmsg_vector->msg_hdr.msg_iov = iov; 576 mmsg_vector->msg_hdr.msg_iovlen = 1; 577 mmsg_vector->msg_hdr.msg_control = NULL; 578 mmsg_vector->msg_hdr.msg_controllen = 0; 579 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT; 580 mmsg_vector->msg_hdr.msg_name = NULL; 581 mmsg_vector->msg_hdr.msg_namelen = 0; 582 if (vp->header_size > 0) { 583 iov->iov_base = kmalloc(header_size, GFP_KERNEL); 584 if (iov->iov_base == NULL) 585 goto out_fail; 586 iov->iov_len = header_size; 587 mmsg_vector->msg_hdr.msg_iovlen = 2; 588 iov++; 589 } 590 iov->iov_base = NULL; 591 iov->iov_len = 0; 592 mmsg_vector++; 593 } 594 spin_lock_init(&result->head_lock); 595 spin_lock_init(&result->tail_lock); 596 result->queue_depth = 0; 597 result->head = 0; 598 result->tail = 0; 599 return result; 600out_skb_fail: 601 kfree(result->mmsg_vector); 602out_mmsg_fail: 603 kfree(result); 604 return NULL; 605out_fail: 606 destroy_queue(result); 607 return NULL; 608} 609 610/* 611 * We do not use the RX queue as a proper wraparound queue for now 612 * This is not necessary because the consumption via netif_rx() 613 * happens in-line. While we can try using the return code of 614 * netif_rx() for flow control there are no drivers doing this today. 615 * For this RX specific use we ignore the tail/head locks and 616 * just read into a prepared queue filled with skbuffs. 617 */ 618 619static struct sk_buff *prep_skb( 620 struct vector_private *vp, 621 struct user_msghdr *msg) 622{ 623 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN; 624 struct sk_buff *result; 625 int iov_index = 0, len; 626 struct iovec *iov = msg->msg_iov; 627 int err, nr_frags, frag; 628 skb_frag_t *skb_frag; 629 630 if (vp->req_size <= linear) 631 len = linear; 632 else 633 len = vp->req_size; 634 result = alloc_skb_with_frags( 635 linear, 636 len - vp->max_packet, 637 3, 638 &err, 639 GFP_ATOMIC 640 ); 641 if (vp->header_size > 0) 642 iov_index++; 643 if (result == NULL) { 644 iov[iov_index].iov_base = NULL; 645 iov[iov_index].iov_len = 0; 646 goto done; 647 } 648 skb_reserve(result, vp->headroom); 649 result->dev = vp->dev; 650 skb_put(result, vp->max_packet); 651 result->data_len = len - vp->max_packet; 652 result->len += len - vp->max_packet; 653 skb_reset_mac_header(result); 654 result->ip_summed = CHECKSUM_NONE; 655 iov[iov_index].iov_base = result->data; 656 iov[iov_index].iov_len = vp->max_packet; 657 iov_index++; 658 659 nr_frags = skb_shinfo(result)->nr_frags; 660 for (frag = 0; frag < nr_frags; frag++) { 661 skb_frag = &skb_shinfo(result)->frags[frag]; 662 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag); 663 if (iov[iov_index].iov_base != NULL) 664 iov[iov_index].iov_len = skb_frag_size(skb_frag); 665 else 666 iov[iov_index].iov_len = 0; 667 iov_index++; 668 } 669done: 670 msg->msg_iovlen = iov_index; 671 return result; 672} 673 674 675/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/ 676 677static void prep_queue_for_rx(struct vector_queue *qi) 678{ 679 struct vector_private *vp = netdev_priv(qi->dev); 680 struct mmsghdr *mmsg_vector = qi->mmsg_vector; 681 void **skbuff_vector = qi->skbuff_vector; 682 int i; 683 684 if (qi->queue_depth == 0) 685 return; 686 for (i = 0; i < qi->queue_depth; i++) { 687 /* it is OK if allocation fails - recvmmsg with NULL data in 688 * iov argument still performs an RX, just drops the packet 689 * This allows us stop faffing around with a "drop buffer" 690 */ 691 692 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr); 693 skbuff_vector++; 694 mmsg_vector++; 695 } 696 qi->queue_depth = 0; 697} 698 699static struct vector_device *find_device(int n) 700{ 701 struct vector_device *device; 702 struct list_head *ele; 703 704 spin_lock(&vector_devices_lock); 705 list_for_each(ele, &vector_devices) { 706 device = list_entry(ele, struct vector_device, list); 707 if (device->unit == n) 708 goto out; 709 } 710 device = NULL; 711 out: 712 spin_unlock(&vector_devices_lock); 713 return device; 714} 715 716static int vector_parse(char *str, int *index_out, char **str_out, 717 char **error_out) 718{ 719 int n, len, err; 720 char *start = str; 721 722 len = strlen(str); 723 724 while ((*str != ':') && (strlen(str) > 1)) 725 str++; 726 if (*str != ':') { 727 *error_out = "Expected ':' after device number"; 728 return -EINVAL; 729 } 730 *str = '\0'; 731 732 err = kstrtouint(start, 0, &n); 733 if (err < 0) { 734 *error_out = "Bad device number"; 735 return err; 736 } 737 738 str++; 739 if (find_device(n)) { 740 *error_out = "Device already configured"; 741 return -EINVAL; 742 } 743 744 *index_out = n; 745 *str_out = str; 746 return 0; 747} 748 749static int vector_config(char *str, char **error_out) 750{ 751 int err, n; 752 char *params; 753 struct arglist *parsed; 754 755 err = vector_parse(str, &n, ¶ms, error_out); 756 if (err != 0) 757 return err; 758 759 /* This string is broken up and the pieces used by the underlying 760 * driver. We should copy it to make sure things do not go wrong 761 * later. 762 */ 763 764 params = kstrdup(params, GFP_KERNEL); 765 if (params == NULL) { 766 *error_out = "vector_config failed to strdup string"; 767 return -ENOMEM; 768 } 769 770 parsed = uml_parse_vector_ifspec(params); 771 772 if (parsed == NULL) { 773 *error_out = "vector_config failed to parse parameters"; 774 kfree(params); 775 return -EINVAL; 776 } 777 778 vector_eth_configure(n, parsed); 779 return 0; 780} 781 782static int vector_id(char **str, int *start_out, int *end_out) 783{ 784 char *end; 785 int n; 786 787 n = simple_strtoul(*str, &end, 0); 788 if ((*end != '\0') || (end == *str)) 789 return -1; 790 791 *start_out = n; 792 *end_out = n; 793 *str = end; 794 return n; 795} 796 797static int vector_remove(int n, char **error_out) 798{ 799 struct vector_device *vec_d; 800 struct net_device *dev; 801 struct vector_private *vp; 802 803 vec_d = find_device(n); 804 if (vec_d == NULL) 805 return -ENODEV; 806 dev = vec_d->dev; 807 vp = netdev_priv(dev); 808 if (vp->fds != NULL) 809 return -EBUSY; 810 unregister_netdev(dev); 811 platform_device_unregister(&vec_d->pdev); 812 return 0; 813} 814 815/* 816 * There is no shared per-transport initialization code, so 817 * we will just initialize each interface one by one and 818 * add them to a list 819 */ 820 821static struct platform_driver uml_net_driver = { 822 .driver = { 823 .name = DRIVER_NAME, 824 }, 825}; 826 827 828static void vector_device_release(struct device *dev) 829{ 830 struct vector_device *device = dev_get_drvdata(dev); 831 struct net_device *netdev = device->dev; 832 833 list_del(&device->list); 834 kfree(device); 835 free_netdev(netdev); 836} 837 838/* Bog standard recv using recvmsg - not used normally unless the user 839 * explicitly specifies not to use recvmmsg vector RX. 840 */ 841 842static int vector_legacy_rx(struct vector_private *vp) 843{ 844 int pkt_len; 845 struct user_msghdr hdr; 846 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */ 847 int iovpos = 0; 848 struct sk_buff *skb; 849 int header_check; 850 851 hdr.msg_name = NULL; 852 hdr.msg_namelen = 0; 853 hdr.msg_iov = (struct iovec *) &iov; 854 hdr.msg_control = NULL; 855 hdr.msg_controllen = 0; 856 hdr.msg_flags = 0; 857 858 if (vp->header_size > 0) { 859 iov[0].iov_base = vp->header_rxbuffer; 860 iov[0].iov_len = vp->header_size; 861 } 862 863 skb = prep_skb(vp, &hdr); 864 865 if (skb == NULL) { 866 /* Read a packet into drop_buffer and don't do 867 * anything with it. 868 */ 869 iov[iovpos].iov_base = drop_buffer; 870 iov[iovpos].iov_len = DROP_BUFFER_SIZE; 871 hdr.msg_iovlen = 1; 872 vp->dev->stats.rx_dropped++; 873 } 874 875 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0); 876 if (pkt_len < 0) { 877 vp->in_error = true; 878 return pkt_len; 879 } 880 881 if (skb != NULL) { 882 if (pkt_len > vp->header_size) { 883 if (vp->header_size > 0) { 884 header_check = vp->verify_header( 885 vp->header_rxbuffer, skb, vp); 886 if (header_check < 0) { 887 dev_kfree_skb_irq(skb); 888 vp->dev->stats.rx_dropped++; 889 vp->estats.rx_encaps_errors++; 890 return 0; 891 } 892 if (header_check > 0) { 893 vp->estats.rx_csum_offload_good++; 894 skb->ip_summed = CHECKSUM_UNNECESSARY; 895 } 896 } 897 pskb_trim(skb, pkt_len - vp->rx_header_size); 898 skb->protocol = eth_type_trans(skb, skb->dev); 899 vp->dev->stats.rx_bytes += skb->len; 900 vp->dev->stats.rx_packets++; 901 netif_rx(skb); 902 } else { 903 dev_kfree_skb_irq(skb); 904 } 905 } 906 return pkt_len; 907} 908 909/* 910 * Packet at a time TX which falls back to vector TX if the 911 * underlying transport is busy. 912 */ 913 914 915 916static int writev_tx(struct vector_private *vp, struct sk_buff *skb) 917{ 918 struct iovec iov[3 + MAX_IOV_SIZE]; 919 int iov_count, pkt_len = 0; 920 921 iov[0].iov_base = vp->header_txbuffer; 922 iov_count = prep_msg(vp, skb, (struct iovec *) &iov); 923 924 if (iov_count < 1) 925 goto drop; 926 927 pkt_len = uml_vector_writev( 928 vp->fds->tx_fd, 929 (struct iovec *) &iov, 930 iov_count 931 ); 932 933 if (pkt_len < 0) 934 goto drop; 935 936 netif_trans_update(vp->dev); 937 netif_wake_queue(vp->dev); 938 939 if (pkt_len > 0) { 940 vp->dev->stats.tx_bytes += skb->len; 941 vp->dev->stats.tx_packets++; 942 } else { 943 vp->dev->stats.tx_dropped++; 944 } 945 consume_skb(skb); 946 return pkt_len; 947drop: 948 vp->dev->stats.tx_dropped++; 949 consume_skb(skb); 950 if (pkt_len < 0) 951 vp->in_error = true; 952 return pkt_len; 953} 954 955/* 956 * Receive as many messages as we can in one call using the special 957 * mmsg vector matched to an skb vector which we prepared earlier. 958 */ 959 960static int vector_mmsg_rx(struct vector_private *vp) 961{ 962 int packet_count, i; 963 struct vector_queue *qi = vp->rx_queue; 964 struct sk_buff *skb; 965 struct mmsghdr *mmsg_vector = qi->mmsg_vector; 966 void **skbuff_vector = qi->skbuff_vector; 967 int header_check; 968 969 /* Refresh the vector and make sure it is with new skbs and the 970 * iovs are updated to point to them. 971 */ 972 973 prep_queue_for_rx(qi); 974 975 /* Fire the Lazy Gun - get as many packets as we can in one go. */ 976 977 packet_count = uml_vector_recvmmsg( 978 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0); 979 980 if (packet_count < 0) 981 vp->in_error = true; 982 983 if (packet_count <= 0) 984 return packet_count; 985 986 /* We treat packet processing as enqueue, buffer refresh as dequeue 987 * The queue_depth tells us how many buffers have been used and how 988 * many do we need to prep the next time prep_queue_for_rx() is called. 989 */ 990 991 qi->queue_depth = packet_count; 992 993 for (i = 0; i < packet_count; i++) { 994 skb = (*skbuff_vector); 995 if (mmsg_vector->msg_len > vp->header_size) { 996 if (vp->header_size > 0) { 997 header_check = vp->verify_header( 998 mmsg_vector->msg_hdr.msg_iov->iov_base, 999 skb, 1000 vp 1001 ); 1002 if (header_check < 0) { 1003 /* Overlay header failed to verify - discard. 1004 * We can actually keep this skb and reuse it, 1005 * but that will make the prep logic too 1006 * complex. 1007 */ 1008 dev_kfree_skb_irq(skb); 1009 vp->estats.rx_encaps_errors++; 1010 continue; 1011 } 1012 if (header_check > 0) { 1013 vp->estats.rx_csum_offload_good++; 1014 skb->ip_summed = CHECKSUM_UNNECESSARY; 1015 } 1016 } 1017 pskb_trim(skb, 1018 mmsg_vector->msg_len - vp->rx_header_size); 1019 skb->protocol = eth_type_trans(skb, skb->dev); 1020 /* 1021 * We do not need to lock on updating stats here 1022 * The interrupt loop is non-reentrant. 1023 */ 1024 vp->dev->stats.rx_bytes += skb->len; 1025 vp->dev->stats.rx_packets++; 1026 netif_rx(skb); 1027 } else { 1028 /* Overlay header too short to do anything - discard. 1029 * We can actually keep this skb and reuse it, 1030 * but that will make the prep logic too complex. 1031 */ 1032 if (skb != NULL) 1033 dev_kfree_skb_irq(skb); 1034 } 1035 (*skbuff_vector) = NULL; 1036 /* Move to the next buffer element */ 1037 mmsg_vector++; 1038 skbuff_vector++; 1039 } 1040 if (packet_count > 0) { 1041 if (vp->estats.rx_queue_max < packet_count) 1042 vp->estats.rx_queue_max = packet_count; 1043 vp->estats.rx_queue_running_average = 1044 (vp->estats.rx_queue_running_average + packet_count) >> 1; 1045 } 1046 return packet_count; 1047} 1048 1049static void vector_rx(struct vector_private *vp) 1050{ 1051 int err; 1052 int iter = 0; 1053 1054 if ((vp->options & VECTOR_RX) > 0) 1055 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS)) 1056 iter++; 1057 else 1058 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS)) 1059 iter++; 1060 if ((err != 0) && net_ratelimit()) 1061 netdev_err(vp->dev, "vector_rx: error(%d)\n", err); 1062 if (iter == MAX_ITERATIONS) 1063 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n"); 1064} 1065 1066static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev) 1067{ 1068 struct vector_private *vp = netdev_priv(dev); 1069 int queue_depth = 0; 1070 1071 if (vp->in_error) { 1072 deactivate_fd(vp->fds->rx_fd, vp->rx_irq); 1073 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0)) 1074 deactivate_fd(vp->fds->tx_fd, vp->tx_irq); 1075 return NETDEV_TX_BUSY; 1076 } 1077 1078 if ((vp->options & VECTOR_TX) == 0) { 1079 writev_tx(vp, skb); 1080 return NETDEV_TX_OK; 1081 } 1082 1083 /* We do BQL only in the vector path, no point doing it in 1084 * packet at a time mode as there is no device queue 1085 */ 1086 1087 netdev_sent_queue(vp->dev, skb->len); 1088 queue_depth = vector_enqueue(vp->tx_queue, skb); 1089 1090 /* if the device queue is full, stop the upper layers and 1091 * flush it. 1092 */ 1093 1094 if (queue_depth >= vp->tx_queue->max_depth - 1) { 1095 vp->estats.tx_kicks++; 1096 netif_stop_queue(dev); 1097 vector_send(vp->tx_queue); 1098 return NETDEV_TX_OK; 1099 } 1100 if (netdev_xmit_more()) { 1101 mod_timer(&vp->tl, vp->coalesce); 1102 return NETDEV_TX_OK; 1103 } 1104 if (skb->len < TX_SMALL_PACKET) { 1105 vp->estats.tx_kicks++; 1106 vector_send(vp->tx_queue); 1107 } else 1108 tasklet_schedule(&vp->tx_poll); 1109 return NETDEV_TX_OK; 1110} 1111 1112static irqreturn_t vector_rx_interrupt(int irq, void *dev_id) 1113{ 1114 struct net_device *dev = dev_id; 1115 struct vector_private *vp = netdev_priv(dev); 1116 1117 if (!netif_running(dev)) 1118 return IRQ_NONE; 1119 vector_rx(vp); 1120 return IRQ_HANDLED; 1121 1122} 1123 1124static irqreturn_t vector_tx_interrupt(int irq, void *dev_id) 1125{ 1126 struct net_device *dev = dev_id; 1127 struct vector_private *vp = netdev_priv(dev); 1128 1129 if (!netif_running(dev)) 1130 return IRQ_NONE; 1131 /* We need to pay attention to it only if we got 1132 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise 1133 * we ignore it. In the future, it may be worth 1134 * it to improve the IRQ controller a bit to make 1135 * tweaking the IRQ mask less costly 1136 */ 1137 1138 if (vp->in_write_poll) 1139 tasklet_schedule(&vp->tx_poll); 1140 return IRQ_HANDLED; 1141 1142} 1143 1144static int irq_rr; 1145 1146static int vector_net_close(struct net_device *dev) 1147{ 1148 struct vector_private *vp = netdev_priv(dev); 1149 unsigned long flags; 1150 1151 netif_stop_queue(dev); 1152 del_timer(&vp->tl); 1153 1154 if (vp->fds == NULL) 1155 return 0; 1156 1157 /* Disable and free all IRQS */ 1158 if (vp->rx_irq > 0) { 1159 um_free_irq(vp->rx_irq, dev); 1160 vp->rx_irq = 0; 1161 } 1162 if (vp->tx_irq > 0) { 1163 um_free_irq(vp->tx_irq, dev); 1164 vp->tx_irq = 0; 1165 } 1166 tasklet_kill(&vp->tx_poll); 1167 if (vp->fds->rx_fd > 0) { 1168 if (vp->bpf) 1169 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf); 1170 os_close_file(vp->fds->rx_fd); 1171 vp->fds->rx_fd = -1; 1172 } 1173 if (vp->fds->tx_fd > 0) { 1174 os_close_file(vp->fds->tx_fd); 1175 vp->fds->tx_fd = -1; 1176 } 1177 if (vp->bpf != NULL) 1178 kfree(vp->bpf->filter); 1179 kfree(vp->bpf); 1180 vp->bpf = NULL; 1181 kfree(vp->fds->remote_addr); 1182 kfree(vp->transport_data); 1183 kfree(vp->header_rxbuffer); 1184 kfree(vp->header_txbuffer); 1185 if (vp->rx_queue != NULL) 1186 destroy_queue(vp->rx_queue); 1187 if (vp->tx_queue != NULL) 1188 destroy_queue(vp->tx_queue); 1189 kfree(vp->fds); 1190 vp->fds = NULL; 1191 spin_lock_irqsave(&vp->lock, flags); 1192 vp->opened = false; 1193 vp->in_error = false; 1194 spin_unlock_irqrestore(&vp->lock, flags); 1195 return 0; 1196} 1197 1198/* TX tasklet */ 1199 1200static void vector_tx_poll(unsigned long data) 1201{ 1202 struct vector_private *vp = (struct vector_private *)data; 1203 1204 vp->estats.tx_kicks++; 1205 vector_send(vp->tx_queue); 1206} 1207static void vector_reset_tx(struct work_struct *work) 1208{ 1209 struct vector_private *vp = 1210 container_of(work, struct vector_private, reset_tx); 1211 netdev_reset_queue(vp->dev); 1212 netif_start_queue(vp->dev); 1213 netif_wake_queue(vp->dev); 1214} 1215 1216static int vector_net_open(struct net_device *dev) 1217{ 1218 struct vector_private *vp = netdev_priv(dev); 1219 unsigned long flags; 1220 int err = -EINVAL; 1221 struct vector_device *vdevice; 1222 1223 spin_lock_irqsave(&vp->lock, flags); 1224 if (vp->opened) { 1225 spin_unlock_irqrestore(&vp->lock, flags); 1226 return -ENXIO; 1227 } 1228 vp->opened = true; 1229 spin_unlock_irqrestore(&vp->lock, flags); 1230 1231 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed)); 1232 1233 vp->fds = uml_vector_user_open(vp->unit, vp->parsed); 1234 1235 if (vp->fds == NULL) 1236 goto out_close; 1237 1238 if (build_transport_data(vp) < 0) 1239 goto out_close; 1240 1241 if ((vp->options & VECTOR_RX) > 0) { 1242 vp->rx_queue = create_queue( 1243 vp, 1244 get_depth(vp->parsed), 1245 vp->rx_header_size, 1246 MAX_IOV_SIZE 1247 ); 1248 vp->rx_queue->queue_depth = get_depth(vp->parsed); 1249 } else { 1250 vp->header_rxbuffer = kmalloc( 1251 vp->rx_header_size, 1252 GFP_KERNEL 1253 ); 1254 if (vp->header_rxbuffer == NULL) 1255 goto out_close; 1256 } 1257 if ((vp->options & VECTOR_TX) > 0) { 1258 vp->tx_queue = create_queue( 1259 vp, 1260 get_depth(vp->parsed), 1261 vp->header_size, 1262 MAX_IOV_SIZE 1263 ); 1264 } else { 1265 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL); 1266 if (vp->header_txbuffer == NULL) 1267 goto out_close; 1268 } 1269 1270 /* READ IRQ */ 1271 err = um_request_irq( 1272 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd, 1273 IRQ_READ, vector_rx_interrupt, 1274 IRQF_SHARED, dev->name, dev); 1275 if (err != 0) { 1276 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err); 1277 err = -ENETUNREACH; 1278 goto out_close; 1279 } 1280 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ; 1281 dev->irq = irq_rr + VECTOR_BASE_IRQ; 1282 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE; 1283 1284 /* WRITE IRQ - we need it only if we have vector TX */ 1285 if ((vp->options & VECTOR_TX) > 0) { 1286 err = um_request_irq( 1287 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd, 1288 IRQ_WRITE, vector_tx_interrupt, 1289 IRQF_SHARED, dev->name, dev); 1290 if (err != 0) { 1291 netdev_err(dev, 1292 "vector_open: failed to get tx irq(%d)\n", err); 1293 err = -ENETUNREACH; 1294 goto out_close; 1295 } 1296 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ; 1297 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE; 1298 } 1299 1300 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) { 1301 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd)) 1302 vp->options |= VECTOR_BPF; 1303 } 1304 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL)) 1305 vp->bpf = uml_vector_default_bpf(dev->dev_addr); 1306 1307 if (vp->bpf != NULL) 1308 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf); 1309 1310 netif_start_queue(dev); 1311 1312 /* clear buffer - it can happen that the host side of the interface 1313 * is full when we get here. In this case, new data is never queued, 1314 * SIGIOs never arrive, and the net never works. 1315 */ 1316 1317 vector_rx(vp); 1318 1319 vector_reset_stats(vp); 1320 vdevice = find_device(vp->unit); 1321 vdevice->opened = 1; 1322 1323 if ((vp->options & VECTOR_TX) != 0) 1324 add_timer(&vp->tl); 1325 return 0; 1326out_close: 1327 vector_net_close(dev); 1328 return err; 1329} 1330 1331 1332static void vector_net_set_multicast_list(struct net_device *dev) 1333{ 1334 /* TODO: - we can do some BPF games here */ 1335 return; 1336} 1337 1338static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue) 1339{ 1340 struct vector_private *vp = netdev_priv(dev); 1341 1342 vp->estats.tx_timeout_count++; 1343 netif_trans_update(dev); 1344 schedule_work(&vp->reset_tx); 1345} 1346 1347static netdev_features_t vector_fix_features(struct net_device *dev, 1348 netdev_features_t features) 1349{ 1350 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 1351 return features; 1352} 1353 1354static int vector_set_features(struct net_device *dev, 1355 netdev_features_t features) 1356{ 1357 struct vector_private *vp = netdev_priv(dev); 1358 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is 1359 * no way to negotiate it on raw sockets, so we can change 1360 * only our side. 1361 */ 1362 if (features & NETIF_F_GRO) 1363 /* All new frame buffers will be GRO-sized */ 1364 vp->req_size = 65536; 1365 else 1366 /* All new frame buffers will be normal sized */ 1367 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN; 1368 return 0; 1369} 1370 1371#ifdef CONFIG_NET_POLL_CONTROLLER 1372static void vector_net_poll_controller(struct net_device *dev) 1373{ 1374 disable_irq(dev->irq); 1375 vector_rx_interrupt(dev->irq, dev); 1376 enable_irq(dev->irq); 1377} 1378#endif 1379 1380static void vector_net_get_drvinfo(struct net_device *dev, 1381 struct ethtool_drvinfo *info) 1382{ 1383 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver)); 1384} 1385 1386static int vector_net_load_bpf_flash(struct net_device *dev, 1387 struct ethtool_flash *efl) 1388{ 1389 struct vector_private *vp = netdev_priv(dev); 1390 struct vector_device *vdevice; 1391 const struct firmware *fw; 1392 int result = 0; 1393 1394 if (!(vp->options & VECTOR_BPF_FLASH)) { 1395 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data); 1396 return -1; 1397 } 1398 1399 spin_lock(&vp->lock); 1400 1401 if (vp->bpf != NULL) { 1402 if (vp->opened) 1403 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf); 1404 kfree(vp->bpf->filter); 1405 vp->bpf->filter = NULL; 1406 } else { 1407 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC); 1408 if (vp->bpf == NULL) { 1409 netdev_err(dev, "failed to allocate memory for firmware\n"); 1410 goto flash_fail; 1411 } 1412 } 1413 1414 vdevice = find_device(vp->unit); 1415 1416 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev)) 1417 goto flash_fail; 1418 1419 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC); 1420 if (!vp->bpf->filter) 1421 goto free_buffer; 1422 1423 vp->bpf->len = fw->size / sizeof(struct sock_filter); 1424 release_firmware(fw); 1425 1426 if (vp->opened) 1427 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf); 1428 1429 spin_unlock(&vp->lock); 1430 1431 return result; 1432 1433free_buffer: 1434 release_firmware(fw); 1435 1436flash_fail: 1437 spin_unlock(&vp->lock); 1438 if (vp->bpf != NULL) 1439 kfree(vp->bpf->filter); 1440 kfree(vp->bpf); 1441 vp->bpf = NULL; 1442 return -1; 1443} 1444 1445static void vector_get_ringparam(struct net_device *netdev, 1446 struct ethtool_ringparam *ring) 1447{ 1448 struct vector_private *vp = netdev_priv(netdev); 1449 1450 ring->rx_max_pending = vp->rx_queue->max_depth; 1451 ring->tx_max_pending = vp->tx_queue->max_depth; 1452 ring->rx_pending = vp->rx_queue->max_depth; 1453 ring->tx_pending = vp->tx_queue->max_depth; 1454} 1455 1456static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf) 1457{ 1458 switch (stringset) { 1459 case ETH_SS_TEST: 1460 *buf = '\0'; 1461 break; 1462 case ETH_SS_STATS: 1463 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys)); 1464 break; 1465 default: 1466 WARN_ON(1); 1467 break; 1468 } 1469} 1470 1471static int vector_get_sset_count(struct net_device *dev, int sset) 1472{ 1473 switch (sset) { 1474 case ETH_SS_TEST: 1475 return 0; 1476 case ETH_SS_STATS: 1477 return VECTOR_NUM_STATS; 1478 default: 1479 return -EOPNOTSUPP; 1480 } 1481} 1482 1483static void vector_get_ethtool_stats(struct net_device *dev, 1484 struct ethtool_stats *estats, 1485 u64 *tmp_stats) 1486{ 1487 struct vector_private *vp = netdev_priv(dev); 1488 1489 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats)); 1490} 1491 1492static int vector_get_coalesce(struct net_device *netdev, 1493 struct ethtool_coalesce *ec) 1494{ 1495 struct vector_private *vp = netdev_priv(netdev); 1496 1497 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ; 1498 return 0; 1499} 1500 1501static int vector_set_coalesce(struct net_device *netdev, 1502 struct ethtool_coalesce *ec) 1503{ 1504 struct vector_private *vp = netdev_priv(netdev); 1505 1506 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000; 1507 if (vp->coalesce == 0) 1508 vp->coalesce = 1; 1509 return 0; 1510} 1511 1512static const struct ethtool_ops vector_net_ethtool_ops = { 1513 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS, 1514 .get_drvinfo = vector_net_get_drvinfo, 1515 .get_link = ethtool_op_get_link, 1516 .get_ts_info = ethtool_op_get_ts_info, 1517 .get_ringparam = vector_get_ringparam, 1518 .get_strings = vector_get_strings, 1519 .get_sset_count = vector_get_sset_count, 1520 .get_ethtool_stats = vector_get_ethtool_stats, 1521 .get_coalesce = vector_get_coalesce, 1522 .set_coalesce = vector_set_coalesce, 1523 .flash_device = vector_net_load_bpf_flash, 1524}; 1525 1526 1527static const struct net_device_ops vector_netdev_ops = { 1528 .ndo_open = vector_net_open, 1529 .ndo_stop = vector_net_close, 1530 .ndo_start_xmit = vector_net_start_xmit, 1531 .ndo_set_rx_mode = vector_net_set_multicast_list, 1532 .ndo_tx_timeout = vector_net_tx_timeout, 1533 .ndo_set_mac_address = eth_mac_addr, 1534 .ndo_validate_addr = eth_validate_addr, 1535 .ndo_fix_features = vector_fix_features, 1536 .ndo_set_features = vector_set_features, 1537#ifdef CONFIG_NET_POLL_CONTROLLER 1538 .ndo_poll_controller = vector_net_poll_controller, 1539#endif 1540}; 1541 1542 1543static void vector_timer_expire(struct timer_list *t) 1544{ 1545 struct vector_private *vp = from_timer(vp, t, tl); 1546 1547 vp->estats.tx_kicks++; 1548 vector_send(vp->tx_queue); 1549} 1550 1551static void vector_eth_configure( 1552 int n, 1553 struct arglist *def 1554 ) 1555{ 1556 struct vector_device *device; 1557 struct net_device *dev; 1558 struct vector_private *vp; 1559 int err; 1560 1561 device = kzalloc(sizeof(*device), GFP_KERNEL); 1562 if (device == NULL) { 1563 printk(KERN_ERR "eth_configure failed to allocate struct " 1564 "vector_device\n"); 1565 return; 1566 } 1567 dev = alloc_etherdev(sizeof(struct vector_private)); 1568 if (dev == NULL) { 1569 printk(KERN_ERR "eth_configure: failed to allocate struct " 1570 "net_device for vec%d\n", n); 1571 goto out_free_device; 1572 } 1573 1574 dev->mtu = get_mtu(def); 1575 1576 INIT_LIST_HEAD(&device->list); 1577 device->unit = n; 1578 1579 /* If this name ends up conflicting with an existing registered 1580 * netdevice, that is OK, register_netdev{,ice}() will notice this 1581 * and fail. 1582 */ 1583 snprintf(dev->name, sizeof(dev->name), "vec%d", n); 1584 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac")); 1585 vp = netdev_priv(dev); 1586 1587 /* sysfs register */ 1588 if (!driver_registered) { 1589 platform_driver_register(¨_net_driver); 1590 driver_registered = 1; 1591 } 1592 device->pdev.id = n; 1593 device->pdev.name = DRIVER_NAME; 1594 device->pdev.dev.release = vector_device_release; 1595 dev_set_drvdata(&device->pdev.dev, device); 1596 if (platform_device_register(&device->pdev)) 1597 goto out_free_netdev; 1598 SET_NETDEV_DEV(dev, &device->pdev.dev); 1599 1600 device->dev = dev; 1601 1602 *vp = ((struct vector_private) 1603 { 1604 .list = LIST_HEAD_INIT(vp->list), 1605 .dev = dev, 1606 .unit = n, 1607 .options = get_transport_options(def), 1608 .rx_irq = 0, 1609 .tx_irq = 0, 1610 .parsed = def, 1611 .max_packet = get_mtu(def) + ETH_HEADER_OTHER, 1612 /* TODO - we need to calculate headroom so that ip header 1613 * is 16 byte aligned all the time 1614 */ 1615 .headroom = get_headroom(def), 1616 .form_header = NULL, 1617 .verify_header = NULL, 1618 .header_rxbuffer = NULL, 1619 .header_txbuffer = NULL, 1620 .header_size = 0, 1621 .rx_header_size = 0, 1622 .rexmit_scheduled = false, 1623 .opened = false, 1624 .transport_data = NULL, 1625 .in_write_poll = false, 1626 .coalesce = 2, 1627 .req_size = get_req_size(def), 1628 .in_error = false, 1629 .bpf = NULL 1630 }); 1631 1632 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST); 1633 tasklet_init(&vp->tx_poll, vector_tx_poll, (unsigned long)vp); 1634 INIT_WORK(&vp->reset_tx, vector_reset_tx); 1635 1636 timer_setup(&vp->tl, vector_timer_expire, 0); 1637 spin_lock_init(&vp->lock); 1638 1639 /* FIXME */ 1640 dev->netdev_ops = &vector_netdev_ops; 1641 dev->ethtool_ops = &vector_net_ethtool_ops; 1642 dev->watchdog_timeo = (HZ >> 1); 1643 /* primary IRQ - fixme */ 1644 dev->irq = 0; /* we will adjust this once opened */ 1645 1646 rtnl_lock(); 1647 err = register_netdevice(dev); 1648 rtnl_unlock(); 1649 if (err) 1650 goto out_undo_user_init; 1651 1652 spin_lock(&vector_devices_lock); 1653 list_add(&device->list, &vector_devices); 1654 spin_unlock(&vector_devices_lock); 1655 1656 return; 1657 1658out_undo_user_init: 1659 return; 1660out_free_netdev: 1661 free_netdev(dev); 1662out_free_device: 1663 kfree(device); 1664} 1665 1666 1667 1668 1669/* 1670 * Invoked late in the init 1671 */ 1672 1673static int __init vector_init(void) 1674{ 1675 struct list_head *ele; 1676 struct vector_cmd_line_arg *def; 1677 struct arglist *parsed; 1678 1679 list_for_each(ele, &vec_cmd_line) { 1680 def = list_entry(ele, struct vector_cmd_line_arg, list); 1681 parsed = uml_parse_vector_ifspec(def->arguments); 1682 if (parsed != NULL) 1683 vector_eth_configure(def->unit, parsed); 1684 } 1685 return 0; 1686} 1687 1688 1689/* Invoked at initial argument parsing, only stores 1690 * arguments until a proper vector_init is called 1691 * later 1692 */ 1693 1694static int __init vector_setup(char *str) 1695{ 1696 char *error; 1697 int n, err; 1698 struct vector_cmd_line_arg *new; 1699 1700 err = vector_parse(str, &n, &str, &error); 1701 if (err) { 1702 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n", 1703 str, error); 1704 return 1; 1705 } 1706 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES); 1707 if (!new) 1708 panic("%s: Failed to allocate %zu bytes\n", __func__, 1709 sizeof(*new)); 1710 INIT_LIST_HEAD(&new->list); 1711 new->unit = n; 1712 new->arguments = str; 1713 list_add_tail(&new->list, &vec_cmd_line); 1714 return 1; 1715} 1716 1717__setup("vec", vector_setup); 1718__uml_help(vector_setup, 1719"vec[0-9]+:<option>=<value>,<option>=<value>\n" 1720" Configure a vector io network device.\n\n" 1721); 1722 1723late_initcall(vector_init); 1724 1725static struct mc_device vector_mc = { 1726 .list = LIST_HEAD_INIT(vector_mc.list), 1727 .name = "vec", 1728 .config = vector_config, 1729 .get_config = NULL, 1730 .id = vector_id, 1731 .remove = vector_remove, 1732}; 1733 1734#ifdef CONFIG_INET 1735static int vector_inetaddr_event( 1736 struct notifier_block *this, 1737 unsigned long event, 1738 void *ptr) 1739{ 1740 return NOTIFY_DONE; 1741} 1742 1743static struct notifier_block vector_inetaddr_notifier = { 1744 .notifier_call = vector_inetaddr_event, 1745}; 1746 1747static void inet_register(void) 1748{ 1749 register_inetaddr_notifier(&vector_inetaddr_notifier); 1750} 1751#else 1752static inline void inet_register(void) 1753{ 1754} 1755#endif 1756 1757static int vector_net_init(void) 1758{ 1759 mconsole_register_dev(&vector_mc); 1760 inet_register(); 1761 return 0; 1762} 1763 1764__initcall(vector_net_init); 1765 1766 1767 1768