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" 49 struct vector_cmd_line_arg { 50 struct list_head list; 51 int unit; 52 char *arguments; 53 }; 54 55 struct 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 63 static LIST_HEAD(vec_cmd_line); 64 65 static DEFINE_SPINLOCK(vector_devices_lock); 66 static LIST_HEAD(vector_devices); 67 68 static int driver_registered; 69 70 static 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 83 static 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) 103 104 static 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 } 119 120 static 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 } 132 133 static char *get_bpf_file(struct arglist *def) 134 { 135 return uml_vector_fetch_arg(def, "bpffile"); 136 } 137 138 static 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 } 149 150 static 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 } 161 162 static 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 } 173 174 static 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 } 188 189 190 static 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 32 233 234 static 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 */ 248 249 static 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 */ 278 279 static 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 } 292 293 static 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; 325 drop: 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 */ 333 334 static 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; 369 drop: 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 } 379 380 static 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 */ 406 407 408 static 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 */ 483 484 static 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 */ 524 static 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; 600 out_skb_fail: 601 kfree(result->mmsg_vector); 602 out_mmsg_fail: 603 kfree(result); 604 return NULL; 605 out_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 */ 618 619 static 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 } 669 done: 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*/ 676 677 static 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 } 698 699 static 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 } 715 716 static 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 } 748 749 static 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 return -EINVAL; 775 } 776 777 vector_eth_configure(n, parsed); 778 return 0; 779 } 780 781 static int vector_id(char **str, int *start_out, int *end_out) 782 { 783 char *end; 784 int n; 785 786 n = simple_strtoul(*str, &end, 0); 787 if ((*end != '\0') || (end == *str)) 788 return -1; 789 790 *start_out = n; 791 *end_out = n; 792 *str = end; 793 return n; 794 } 795 796 static int vector_remove(int n, char **error_out) 797 { 798 struct vector_device *vec_d; 799 struct net_device *dev; 800 struct vector_private *vp; 801 802 vec_d = find_device(n); 803 if (vec_d == NULL) 804 return -ENODEV; 805 dev = vec_d->dev; 806 vp = netdev_priv(dev); 807 if (vp->fds != NULL) 808 return -EBUSY; 809 unregister_netdev(dev); 810 platform_device_unregister(&vec_d->pdev); 811 return 0; 812 } 813 814 /* 815 * There is no shared per-transport initialization code, so 816 * we will just initialize each interface one by one and 817 * add them to a list 818 */ 819 820 static struct platform_driver uml_net_driver = { 821 .driver = { 822 .name = DRIVER_NAME, 823 }, 824 }; 825 826 827 static void vector_device_release(struct device *dev) 828 { 829 struct vector_device *device = dev_get_drvdata(dev); 830 struct net_device *netdev = device->dev; 831 832 list_del(&device->list); 833 kfree(device); 834 free_netdev(netdev); 835 } 836 837 /* Bog standard recv using recvmsg - not used normally unless the user 838 * explicitly specifies not to use recvmmsg vector RX. 839 */ 840 841 static int vector_legacy_rx(struct vector_private *vp) 842 { 843 int pkt_len; 844 struct user_msghdr hdr; 845 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */ 846 int iovpos = 0; 847 struct sk_buff *skb; 848 int header_check; 849 850 hdr.msg_name = NULL; 851 hdr.msg_namelen = 0; 852 hdr.msg_iov = (struct iovec *) &iov; 853 hdr.msg_control = NULL; 854 hdr.msg_controllen = 0; 855 hdr.msg_flags = 0; 856 857 if (vp->header_size > 0) { 858 iov[0].iov_base = vp->header_rxbuffer; 859 iov[0].iov_len = vp->header_size; 860 } 861 862 skb = prep_skb(vp, &hdr); 863 864 if (skb == NULL) { 865 /* Read a packet into drop_buffer and don't do 866 * anything with it. 867 */ 868 iov[iovpos].iov_base = drop_buffer; 869 iov[iovpos].iov_len = DROP_BUFFER_SIZE; 870 hdr.msg_iovlen = 1; 871 vp->dev->stats.rx_dropped++; 872 } 873 874 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0); 875 if (pkt_len < 0) { 876 vp->in_error = true; 877 return pkt_len; 878 } 879 880 if (skb != NULL) { 881 if (pkt_len > vp->header_size) { 882 if (vp->header_size > 0) { 883 header_check = vp->verify_header( 884 vp->header_rxbuffer, skb, vp); 885 if (header_check < 0) { 886 dev_kfree_skb_irq(skb); 887 vp->dev->stats.rx_dropped++; 888 vp->estats.rx_encaps_errors++; 889 return 0; 890 } 891 if (header_check > 0) { 892 vp->estats.rx_csum_offload_good++; 893 skb->ip_summed = CHECKSUM_UNNECESSARY; 894 } 895 } 896 pskb_trim(skb, pkt_len - vp->rx_header_size); 897 skb->protocol = eth_type_trans(skb, skb->dev); 898 vp->dev->stats.rx_bytes += skb->len; 899 vp->dev->stats.rx_packets++; 900 netif_rx(skb); 901 } else { 902 dev_kfree_skb_irq(skb); 903 } 904 } 905 return pkt_len; 906 } 907 908 /* 909 * Packet at a time TX which falls back to vector TX if the 910 * underlying transport is busy. 911 */ 912 913 914 915 static int writev_tx(struct vector_private *vp, struct sk_buff *skb) 916 { 917 struct iovec iov[3 + MAX_IOV_SIZE]; 918 int iov_count, pkt_len = 0; 919 920 iov[0].iov_base = vp->header_txbuffer; 921 iov_count = prep_msg(vp, skb, (struct iovec *) &iov); 922 923 if (iov_count < 1) 924 goto drop; 925 926 pkt_len = uml_vector_writev( 927 vp->fds->tx_fd, 928 (struct iovec *) &iov, 929 iov_count 930 ); 931 932 if (pkt_len < 0) 933 goto drop; 934 935 netif_trans_update(vp->dev); 936 netif_wake_queue(vp->dev); 937 938 if (pkt_len > 0) { 939 vp->dev->stats.tx_bytes += skb->len; 940 vp->dev->stats.tx_packets++; 941 } else { 942 vp->dev->stats.tx_dropped++; 943 } 944 consume_skb(skb); 945 return pkt_len; 946 drop: 947 vp->dev->stats.tx_dropped++; 948 consume_skb(skb); 949 if (pkt_len < 0) 950 vp->in_error = true; 951 return pkt_len; 952 } 953 954 /* 955 * Receive as many messages as we can in one call using the special 956 * mmsg vector matched to an skb vector which we prepared earlier. 957 */ 958 959 static int vector_mmsg_rx(struct vector_private *vp) 960 { 961 int packet_count, i; 962 struct vector_queue *qi = vp->rx_queue; 963 struct sk_buff *skb; 964 struct mmsghdr *mmsg_vector = qi->mmsg_vector; 965 void **skbuff_vector = qi->skbuff_vector; 966 int header_check; 967 968 /* Refresh the vector and make sure it is with new skbs and the 969 * iovs are updated to point to them. 970 */ 971 972 prep_queue_for_rx(qi); 973 974 /* Fire the Lazy Gun - get as many packets as we can in one go. */ 975 976 packet_count = uml_vector_recvmmsg( 977 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0); 978 979 if (packet_count < 0) 980 vp->in_error = true; 981 982 if (packet_count <= 0) 983 return packet_count; 984 985 /* We treat packet processing as enqueue, buffer refresh as dequeue 986 * The queue_depth tells us how many buffers have been used and how 987 * many do we need to prep the next time prep_queue_for_rx() is called. 988 */ 989 990 qi->queue_depth = packet_count; 991 992 for (i = 0; i < packet_count; i++) { 993 skb = (*skbuff_vector); 994 if (mmsg_vector->msg_len > vp->header_size) { 995 if (vp->header_size > 0) { 996 header_check = vp->verify_header( 997 mmsg_vector->msg_hdr.msg_iov->iov_base, 998 skb, 999 vp 1000 ); 1001 if (header_check < 0) { 1002 /* Overlay header failed to verify - discard. 1003 * We can actually keep this skb and reuse it, 1004 * but that will make the prep logic too 1005 * complex. 1006 */ 1007 dev_kfree_skb_irq(skb); 1008 vp->estats.rx_encaps_errors++; 1009 continue; 1010 } 1011 if (header_check > 0) { 1012 vp->estats.rx_csum_offload_good++; 1013 skb->ip_summed = CHECKSUM_UNNECESSARY; 1014 } 1015 } 1016 pskb_trim(skb, 1017 mmsg_vector->msg_len - vp->rx_header_size); 1018 skb->protocol = eth_type_trans(skb, skb->dev); 1019 /* 1020 * We do not need to lock on updating stats here 1021 * The interrupt loop is non-reentrant. 1022 */ 1023 vp->dev->stats.rx_bytes += skb->len; 1024 vp->dev->stats.rx_packets++; 1025 netif_rx(skb); 1026 } else { 1027 /* Overlay header too short to do anything - discard. 1028 * We can actually keep this skb and reuse it, 1029 * but that will make the prep logic too complex. 1030 */ 1031 if (skb != NULL) 1032 dev_kfree_skb_irq(skb); 1033 } 1034 (*skbuff_vector) = NULL; 1035 /* Move to the next buffer element */ 1036 mmsg_vector++; 1037 skbuff_vector++; 1038 } 1039 if (packet_count > 0) { 1040 if (vp->estats.rx_queue_max < packet_count) 1041 vp->estats.rx_queue_max = packet_count; 1042 vp->estats.rx_queue_running_average = 1043 (vp->estats.rx_queue_running_average + packet_count) >> 1; 1044 } 1045 return packet_count; 1046 } 1047 1048 static void vector_rx(struct vector_private *vp) 1049 { 1050 int err; 1051 int iter = 0; 1052 1053 if ((vp->options & VECTOR_RX) > 0) 1054 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS)) 1055 iter++; 1056 else 1057 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS)) 1058 iter++; 1059 if ((err != 0) && net_ratelimit()) 1060 netdev_err(vp->dev, "vector_rx: error(%d)\n", err); 1061 if (iter == MAX_ITERATIONS) 1062 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n"); 1063 } 1064 1065 static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev) 1066 { 1067 struct vector_private *vp = netdev_priv(dev); 1068 int queue_depth = 0; 1069 1070 if (vp->in_error) { 1071 deactivate_fd(vp->fds->rx_fd, vp->rx_irq); 1072 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0)) 1073 deactivate_fd(vp->fds->tx_fd, vp->tx_irq); 1074 return NETDEV_TX_BUSY; 1075 } 1076 1077 if ((vp->options & VECTOR_TX) == 0) { 1078 writev_tx(vp, skb); 1079 return NETDEV_TX_OK; 1080 } 1081 1082 /* We do BQL only in the vector path, no point doing it in 1083 * packet at a time mode as there is no device queue 1084 */ 1085 1086 netdev_sent_queue(vp->dev, skb->len); 1087 queue_depth = vector_enqueue(vp->tx_queue, skb); 1088 1089 /* if the device queue is full, stop the upper layers and 1090 * flush it. 1091 */ 1092 1093 if (queue_depth >= vp->tx_queue->max_depth - 1) { 1094 vp->estats.tx_kicks++; 1095 netif_stop_queue(dev); 1096 vector_send(vp->tx_queue); 1097 return NETDEV_TX_OK; 1098 } 1099 if (netdev_xmit_more()) { 1100 mod_timer(&vp->tl, vp->coalesce); 1101 return NETDEV_TX_OK; 1102 } 1103 if (skb->len < TX_SMALL_PACKET) { 1104 vp->estats.tx_kicks++; 1105 vector_send(vp->tx_queue); 1106 } else 1107 tasklet_schedule(&vp->tx_poll); 1108 return NETDEV_TX_OK; 1109 } 1110 1111 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id) 1112 { 1113 struct net_device *dev = dev_id; 1114 struct vector_private *vp = netdev_priv(dev); 1115 1116 if (!netif_running(dev)) 1117 return IRQ_NONE; 1118 vector_rx(vp); 1119 return IRQ_HANDLED; 1120 1121 } 1122 1123 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id) 1124 { 1125 struct net_device *dev = dev_id; 1126 struct vector_private *vp = netdev_priv(dev); 1127 1128 if (!netif_running(dev)) 1129 return IRQ_NONE; 1130 /* We need to pay attention to it only if we got 1131 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise 1132 * we ignore it. In the future, it may be worth 1133 * it to improve the IRQ controller a bit to make 1134 * tweaking the IRQ mask less costly 1135 */ 1136 1137 if (vp->in_write_poll) 1138 tasklet_schedule(&vp->tx_poll); 1139 return IRQ_HANDLED; 1140 1141 } 1142 1143 static int irq_rr; 1144 1145 static int vector_net_close(struct net_device *dev) 1146 { 1147 struct vector_private *vp = netdev_priv(dev); 1148 unsigned long flags; 1149 1150 netif_stop_queue(dev); 1151 del_timer(&vp->tl); 1152 1153 if (vp->fds == NULL) 1154 return 0; 1155 1156 /* Disable and free all IRQS */ 1157 if (vp->rx_irq > 0) { 1158 um_free_irq(vp->rx_irq, dev); 1159 vp->rx_irq = 0; 1160 } 1161 if (vp->tx_irq > 0) { 1162 um_free_irq(vp->tx_irq, dev); 1163 vp->tx_irq = 0; 1164 } 1165 tasklet_kill(&vp->tx_poll); 1166 if (vp->fds->rx_fd > 0) { 1167 if (vp->bpf) 1168 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf); 1169 os_close_file(vp->fds->rx_fd); 1170 vp->fds->rx_fd = -1; 1171 } 1172 if (vp->fds->tx_fd > 0) { 1173 os_close_file(vp->fds->tx_fd); 1174 vp->fds->tx_fd = -1; 1175 } 1176 if (vp->bpf != NULL) 1177 kfree(vp->bpf->filter); 1178 kfree(vp->bpf); 1179 vp->bpf = NULL; 1180 kfree(vp->fds->remote_addr); 1181 kfree(vp->transport_data); 1182 kfree(vp->header_rxbuffer); 1183 kfree(vp->header_txbuffer); 1184 if (vp->rx_queue != NULL) 1185 destroy_queue(vp->rx_queue); 1186 if (vp->tx_queue != NULL) 1187 destroy_queue(vp->tx_queue); 1188 kfree(vp->fds); 1189 vp->fds = NULL; 1190 spin_lock_irqsave(&vp->lock, flags); 1191 vp->opened = false; 1192 vp->in_error = false; 1193 spin_unlock_irqrestore(&vp->lock, flags); 1194 return 0; 1195 } 1196 1197 /* TX tasklet */ 1198 1199 static void vector_tx_poll(struct tasklet_struct *t) 1200 { 1201 struct vector_private *vp = from_tasklet(vp, t, tx_poll); 1202 1203 vp->estats.tx_kicks++; 1204 vector_send(vp->tx_queue); 1205 } 1206 static void vector_reset_tx(struct work_struct *work) 1207 { 1208 struct vector_private *vp = 1209 container_of(work, struct vector_private, reset_tx); 1210 netdev_reset_queue(vp->dev); 1211 netif_start_queue(vp->dev); 1212 netif_wake_queue(vp->dev); 1213 } 1214 1215 static int vector_net_open(struct net_device *dev) 1216 { 1217 struct vector_private *vp = netdev_priv(dev); 1218 unsigned long flags; 1219 int err = -EINVAL; 1220 struct vector_device *vdevice; 1221 1222 spin_lock_irqsave(&vp->lock, flags); 1223 if (vp->opened) { 1224 spin_unlock_irqrestore(&vp->lock, flags); 1225 return -ENXIO; 1226 } 1227 vp->opened = true; 1228 spin_unlock_irqrestore(&vp->lock, flags); 1229 1230 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed)); 1231 1232 vp->fds = uml_vector_user_open(vp->unit, vp->parsed); 1233 1234 if (vp->fds == NULL) 1235 goto out_close; 1236 1237 if (build_transport_data(vp) < 0) 1238 goto out_close; 1239 1240 if ((vp->options & VECTOR_RX) > 0) { 1241 vp->rx_queue = create_queue( 1242 vp, 1243 get_depth(vp->parsed), 1244 vp->rx_header_size, 1245 MAX_IOV_SIZE 1246 ); 1247 vp->rx_queue->queue_depth = get_depth(vp->parsed); 1248 } else { 1249 vp->header_rxbuffer = kmalloc( 1250 vp->rx_header_size, 1251 GFP_KERNEL 1252 ); 1253 if (vp->header_rxbuffer == NULL) 1254 goto out_close; 1255 } 1256 if ((vp->options & VECTOR_TX) > 0) { 1257 vp->tx_queue = create_queue( 1258 vp, 1259 get_depth(vp->parsed), 1260 vp->header_size, 1261 MAX_IOV_SIZE 1262 ); 1263 } else { 1264 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL); 1265 if (vp->header_txbuffer == NULL) 1266 goto out_close; 1267 } 1268 1269 /* READ IRQ */ 1270 err = um_request_irq( 1271 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd, 1272 IRQ_READ, vector_rx_interrupt, 1273 IRQF_SHARED, dev->name, dev); 1274 if (err < 0) { 1275 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err); 1276 err = -ENETUNREACH; 1277 goto out_close; 1278 } 1279 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ; 1280 dev->irq = irq_rr + VECTOR_BASE_IRQ; 1281 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE; 1282 1283 /* WRITE IRQ - we need it only if we have vector TX */ 1284 if ((vp->options & VECTOR_TX) > 0) { 1285 err = um_request_irq( 1286 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd, 1287 IRQ_WRITE, vector_tx_interrupt, 1288 IRQF_SHARED, dev->name, dev); 1289 if (err < 0) { 1290 netdev_err(dev, 1291 "vector_open: failed to get tx irq(%d)\n", err); 1292 err = -ENETUNREACH; 1293 goto out_close; 1294 } 1295 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ; 1296 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE; 1297 } 1298 1299 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) { 1300 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd)) 1301 vp->options |= VECTOR_BPF; 1302 } 1303 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL)) 1304 vp->bpf = uml_vector_default_bpf(dev->dev_addr); 1305 1306 if (vp->bpf != NULL) 1307 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf); 1308 1309 netif_start_queue(dev); 1310 1311 /* clear buffer - it can happen that the host side of the interface 1312 * is full when we get here. In this case, new data is never queued, 1313 * SIGIOs never arrive, and the net never works. 1314 */ 1315 1316 vector_rx(vp); 1317 1318 vector_reset_stats(vp); 1319 vdevice = find_device(vp->unit); 1320 vdevice->opened = 1; 1321 1322 if ((vp->options & VECTOR_TX) != 0) 1323 add_timer(&vp->tl); 1324 return 0; 1325 out_close: 1326 vector_net_close(dev); 1327 return err; 1328 } 1329 1330 1331 static void vector_net_set_multicast_list(struct net_device *dev) 1332 { 1333 /* TODO: - we can do some BPF games here */ 1334 return; 1335 } 1336 1337 static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue) 1338 { 1339 struct vector_private *vp = netdev_priv(dev); 1340 1341 vp->estats.tx_timeout_count++; 1342 netif_trans_update(dev); 1343 schedule_work(&vp->reset_tx); 1344 } 1345 1346 static netdev_features_t vector_fix_features(struct net_device *dev, 1347 netdev_features_t features) 1348 { 1349 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 1350 return features; 1351 } 1352 1353 static int vector_set_features(struct net_device *dev, 1354 netdev_features_t features) 1355 { 1356 struct vector_private *vp = netdev_priv(dev); 1357 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is 1358 * no way to negotiate it on raw sockets, so we can change 1359 * only our side. 1360 */ 1361 if (features & NETIF_F_GRO) 1362 /* All new frame buffers will be GRO-sized */ 1363 vp->req_size = 65536; 1364 else 1365 /* All new frame buffers will be normal sized */ 1366 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN; 1367 return 0; 1368 } 1369 1370 #ifdef CONFIG_NET_POLL_CONTROLLER 1371 static void vector_net_poll_controller(struct net_device *dev) 1372 { 1373 disable_irq(dev->irq); 1374 vector_rx_interrupt(dev->irq, dev); 1375 enable_irq(dev->irq); 1376 } 1377 #endif 1378 1379 static void vector_net_get_drvinfo(struct net_device *dev, 1380 struct ethtool_drvinfo *info) 1381 { 1382 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver)); 1383 } 1384 1385 static int vector_net_load_bpf_flash(struct net_device *dev, 1386 struct ethtool_flash *efl) 1387 { 1388 struct vector_private *vp = netdev_priv(dev); 1389 struct vector_device *vdevice; 1390 const struct firmware *fw; 1391 int result = 0; 1392 1393 if (!(vp->options & VECTOR_BPF_FLASH)) { 1394 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data); 1395 return -1; 1396 } 1397 1398 spin_lock(&vp->lock); 1399 1400 if (vp->bpf != NULL) { 1401 if (vp->opened) 1402 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf); 1403 kfree(vp->bpf->filter); 1404 vp->bpf->filter = NULL; 1405 } else { 1406 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC); 1407 if (vp->bpf == NULL) { 1408 netdev_err(dev, "failed to allocate memory for firmware\n"); 1409 goto flash_fail; 1410 } 1411 } 1412 1413 vdevice = find_device(vp->unit); 1414 1415 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev)) 1416 goto flash_fail; 1417 1418 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC); 1419 if (!vp->bpf->filter) 1420 goto free_buffer; 1421 1422 vp->bpf->len = fw->size / sizeof(struct sock_filter); 1423 release_firmware(fw); 1424 1425 if (vp->opened) 1426 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf); 1427 1428 spin_unlock(&vp->lock); 1429 1430 return result; 1431 1432 free_buffer: 1433 release_firmware(fw); 1434 1435 flash_fail: 1436 spin_unlock(&vp->lock); 1437 if (vp->bpf != NULL) 1438 kfree(vp->bpf->filter); 1439 kfree(vp->bpf); 1440 vp->bpf = NULL; 1441 return -1; 1442 } 1443 1444 static void vector_get_ringparam(struct net_device *netdev, 1445 struct ethtool_ringparam *ring) 1446 { 1447 struct vector_private *vp = netdev_priv(netdev); 1448 1449 ring->rx_max_pending = vp->rx_queue->max_depth; 1450 ring->tx_max_pending = vp->tx_queue->max_depth; 1451 ring->rx_pending = vp->rx_queue->max_depth; 1452 ring->tx_pending = vp->tx_queue->max_depth; 1453 } 1454 1455 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf) 1456 { 1457 switch (stringset) { 1458 case ETH_SS_TEST: 1459 *buf = '\0'; 1460 break; 1461 case ETH_SS_STATS: 1462 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys)); 1463 break; 1464 default: 1465 WARN_ON(1); 1466 break; 1467 } 1468 } 1469 1470 static int vector_get_sset_count(struct net_device *dev, int sset) 1471 { 1472 switch (sset) { 1473 case ETH_SS_TEST: 1474 return 0; 1475 case ETH_SS_STATS: 1476 return VECTOR_NUM_STATS; 1477 default: 1478 return -EOPNOTSUPP; 1479 } 1480 } 1481 1482 static void vector_get_ethtool_stats(struct net_device *dev, 1483 struct ethtool_stats *estats, 1484 u64 *tmp_stats) 1485 { 1486 struct vector_private *vp = netdev_priv(dev); 1487 1488 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats)); 1489 } 1490 1491 static int vector_get_coalesce(struct net_device *netdev, 1492 struct ethtool_coalesce *ec) 1493 { 1494 struct vector_private *vp = netdev_priv(netdev); 1495 1496 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ; 1497 return 0; 1498 } 1499 1500 static int vector_set_coalesce(struct net_device *netdev, 1501 struct ethtool_coalesce *ec) 1502 { 1503 struct vector_private *vp = netdev_priv(netdev); 1504 1505 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000; 1506 if (vp->coalesce == 0) 1507 vp->coalesce = 1; 1508 return 0; 1509 } 1510 1511 static const struct ethtool_ops vector_net_ethtool_ops = { 1512 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS, 1513 .get_drvinfo = vector_net_get_drvinfo, 1514 .get_link = ethtool_op_get_link, 1515 .get_ts_info = ethtool_op_get_ts_info, 1516 .get_ringparam = vector_get_ringparam, 1517 .get_strings = vector_get_strings, 1518 .get_sset_count = vector_get_sset_count, 1519 .get_ethtool_stats = vector_get_ethtool_stats, 1520 .get_coalesce = vector_get_coalesce, 1521 .set_coalesce = vector_set_coalesce, 1522 .flash_device = vector_net_load_bpf_flash, 1523 }; 1524 1525 1526 static const struct net_device_ops vector_netdev_ops = { 1527 .ndo_open = vector_net_open, 1528 .ndo_stop = vector_net_close, 1529 .ndo_start_xmit = vector_net_start_xmit, 1530 .ndo_set_rx_mode = vector_net_set_multicast_list, 1531 .ndo_tx_timeout = vector_net_tx_timeout, 1532 .ndo_set_mac_address = eth_mac_addr, 1533 .ndo_validate_addr = eth_validate_addr, 1534 .ndo_fix_features = vector_fix_features, 1535 .ndo_set_features = vector_set_features, 1536 #ifdef CONFIG_NET_POLL_CONTROLLER 1537 .ndo_poll_controller = vector_net_poll_controller, 1538 #endif 1539 }; 1540 1541 1542 static void vector_timer_expire(struct timer_list *t) 1543 { 1544 struct vector_private *vp = from_timer(vp, t, tl); 1545 1546 vp->estats.tx_kicks++; 1547 vector_send(vp->tx_queue); 1548 } 1549 1550 static void vector_eth_configure( 1551 int n, 1552 struct arglist *def 1553 ) 1554 { 1555 struct vector_device *device; 1556 struct net_device *dev; 1557 struct vector_private *vp; 1558 int err; 1559 1560 device = kzalloc(sizeof(*device), GFP_KERNEL); 1561 if (device == NULL) { 1562 printk(KERN_ERR "eth_configure failed to allocate struct " 1563 "vector_device\n"); 1564 return; 1565 } 1566 dev = alloc_etherdev(sizeof(struct vector_private)); 1567 if (dev == NULL) { 1568 printk(KERN_ERR "eth_configure: failed to allocate struct " 1569 "net_device for vec%d\n", n); 1570 goto out_free_device; 1571 } 1572 1573 dev->mtu = get_mtu(def); 1574 1575 INIT_LIST_HEAD(&device->list); 1576 device->unit = n; 1577 1578 /* If this name ends up conflicting with an existing registered 1579 * netdevice, that is OK, register_netdev{,ice}() will notice this 1580 * and fail. 1581 */ 1582 snprintf(dev->name, sizeof(dev->name), "vec%d", n); 1583 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac")); 1584 vp = netdev_priv(dev); 1585 1586 /* sysfs register */ 1587 if (!driver_registered) { 1588 platform_driver_register(¨_net_driver); 1589 driver_registered = 1; 1590 } 1591 device->pdev.id = n; 1592 device->pdev.name = DRIVER_NAME; 1593 device->pdev.dev.release = vector_device_release; 1594 dev_set_drvdata(&device->pdev.dev, device); 1595 if (platform_device_register(&device->pdev)) 1596 goto out_free_netdev; 1597 SET_NETDEV_DEV(dev, &device->pdev.dev); 1598 1599 device->dev = dev; 1600 1601 *vp = ((struct vector_private) 1602 { 1603 .list = LIST_HEAD_INIT(vp->list), 1604 .dev = dev, 1605 .unit = n, 1606 .options = get_transport_options(def), 1607 .rx_irq = 0, 1608 .tx_irq = 0, 1609 .parsed = def, 1610 .max_packet = get_mtu(def) + ETH_HEADER_OTHER, 1611 /* TODO - we need to calculate headroom so that ip header 1612 * is 16 byte aligned all the time 1613 */ 1614 .headroom = get_headroom(def), 1615 .form_header = NULL, 1616 .verify_header = NULL, 1617 .header_rxbuffer = NULL, 1618 .header_txbuffer = NULL, 1619 .header_size = 0, 1620 .rx_header_size = 0, 1621 .rexmit_scheduled = false, 1622 .opened = false, 1623 .transport_data = NULL, 1624 .in_write_poll = false, 1625 .coalesce = 2, 1626 .req_size = get_req_size(def), 1627 .in_error = false, 1628 .bpf = NULL 1629 }); 1630 1631 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST); 1632 tasklet_setup(&vp->tx_poll, vector_tx_poll); 1633 INIT_WORK(&vp->reset_tx, vector_reset_tx); 1634 1635 timer_setup(&vp->tl, vector_timer_expire, 0); 1636 spin_lock_init(&vp->lock); 1637 1638 /* FIXME */ 1639 dev->netdev_ops = &vector_netdev_ops; 1640 dev->ethtool_ops = &vector_net_ethtool_ops; 1641 dev->watchdog_timeo = (HZ >> 1); 1642 /* primary IRQ - fixme */ 1643 dev->irq = 0; /* we will adjust this once opened */ 1644 1645 rtnl_lock(); 1646 err = register_netdevice(dev); 1647 rtnl_unlock(); 1648 if (err) 1649 goto out_undo_user_init; 1650 1651 spin_lock(&vector_devices_lock); 1652 list_add(&device->list, &vector_devices); 1653 spin_unlock(&vector_devices_lock); 1654 1655 return; 1656 1657 out_undo_user_init: 1658 return; 1659 out_free_netdev: 1660 free_netdev(dev); 1661 out_free_device: 1662 kfree(device); 1663 } 1664 1665 1666 1667 1668 /* 1669 * Invoked late in the init 1670 */ 1671 1672 static int __init vector_init(void) 1673 { 1674 struct list_head *ele; 1675 struct vector_cmd_line_arg *def; 1676 struct arglist *parsed; 1677 1678 list_for_each(ele, &vec_cmd_line) { 1679 def = list_entry(ele, struct vector_cmd_line_arg, list); 1680 parsed = uml_parse_vector_ifspec(def->arguments); 1681 if (parsed != NULL) 1682 vector_eth_configure(def->unit, parsed); 1683 } 1684 return 0; 1685 } 1686 1687 1688 /* Invoked at initial argument parsing, only stores 1689 * arguments until a proper vector_init is called 1690 * later 1691 */ 1692 1693 static int __init vector_setup(char *str) 1694 { 1695 char *error; 1696 int n, err; 1697 struct vector_cmd_line_arg *new; 1698 1699 err = vector_parse(str, &n, &str, &error); 1700 if (err) { 1701 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n", 1702 str, error); 1703 return 1; 1704 } 1705 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES); 1706 if (!new) 1707 panic("%s: Failed to allocate %zu bytes\n", __func__, 1708 sizeof(*new)); 1709 INIT_LIST_HEAD(&new->list); 1710 new->unit = n; 1711 new->arguments = str; 1712 list_add_tail(&new->list, &vec_cmd_line); 1713 return 1; 1714 } 1715 1716 __setup("vec", vector_setup); 1717 __uml_help(vector_setup, 1718 "vec[0-9]+:<option>=<value>,<option>=<value>\n" 1719 " Configure a vector io network device.\n\n" 1720 ); 1721 1722 late_initcall(vector_init); 1723 1724 static struct mc_device vector_mc = { 1725 .list = LIST_HEAD_INIT(vector_mc.list), 1726 .name = "vec", 1727 .config = vector_config, 1728 .get_config = NULL, 1729 .id = vector_id, 1730 .remove = vector_remove, 1731 }; 1732 1733 #ifdef CONFIG_INET 1734 static int vector_inetaddr_event( 1735 struct notifier_block *this, 1736 unsigned long event, 1737 void *ptr) 1738 { 1739 return NOTIFY_DONE; 1740 } 1741 1742 static struct notifier_block vector_inetaddr_notifier = { 1743 .notifier_call = vector_inetaddr_event, 1744 }; 1745 1746 static void inet_register(void) 1747 { 1748 register_inetaddr_notifier(&vector_inetaddr_notifier); 1749 } 1750 #else 1751 static inline void inet_register(void) 1752 { 1753 } 1754 #endif 1755 1756 static int vector_net_init(void) 1757 { 1758 mconsole_register_dev(&vector_mc); 1759 inet_register(); 1760 return 0; 1761 } 1762 1763 __initcall(vector_net_init); 1764 1765 1766 1767