vector_kern.c

   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#define pr_fmt(fmt) "uml-vector: " fmt
  12
  13#include <linux/memblock.h>
  14#include <linux/etherdevice.h>
  15#include <linux/ethtool.h>
  16#include <linux/inetdevice.h>
  17#include <linux/init.h>
  18#include <linux/list.h>
  19#include <linux/netdevice.h>
  20#include <linux/platform_device.h>
  21#include <linux/rtnetlink.h>
  22#include <linux/skbuff.h>
  23#include <linux/slab.h>
  24#include <linux/interrupt.h>
  25#include <linux/firmware.h>
  26#include <linux/fs.h>
  27#include <asm/atomic.h>
  28#include <uapi/linux/filter.h>
  29#include <init.h>
  30#include <irq_kern.h>
  31#include <irq_user.h>
  32#include <os.h>
  33#include "mconsole_kern.h"
  34#include "vector_user.h"
  35#include "vector_kern.h"
  36
  37/*
  38 * Adapted from network devices with the following major changes:
  39 * All transports are static - simplifies the code significantly
  40 * Multiple FDs/IRQs per device
  41 * Vector IO optionally used for read/write, falling back to legacy
  42 * based on configuration and/or availability
  43 * Configuration is no longer positional - L2TPv3 and GRE require up to
  44 * 10 parameters, passing this as positional is not fit for purpose.
  45 * Only socket transports are supported
  46 */
  47
  48
  49#define DRIVER_NAME "uml-vector"
  50struct vector_cmd_line_arg {
  51	struct list_head list;
  52	int unit;
  53	char *arguments;
  54};
  55
  56struct vector_device {
  57	struct list_head list;
  58	struct net_device *dev;
  59	struct platform_device pdev;
  60	int unit;
  61	int opened;
  62};
  63
  64static LIST_HEAD(vec_cmd_line);
  65
  66static DEFINE_SPINLOCK(vector_devices_lock);
  67static LIST_HEAD(vector_devices);
  68
  69static int driver_registered;
  70
  71static void vector_eth_configure(int n, struct arglist *def);
  72static int vector_mmsg_rx(struct vector_private *vp, int budget);
  73
  74/* Argument accessors to set variables (and/or set default values)
  75 * mtu, buffer sizing, default headroom, etc
  76 */
  77
  78#define DEFAULT_HEADROOM 2
  79#define SAFETY_MARGIN 32
  80#define DEFAULT_VECTOR_SIZE 64
  81#define TX_SMALL_PACKET 128
  82#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
  83
  84static const struct {
  85	const char string[ETH_GSTRING_LEN];
  86} ethtool_stats_keys[] = {
  87	{ "rx_queue_max" },
  88	{ "rx_queue_running_average" },
  89	{ "tx_queue_max" },
  90	{ "tx_queue_running_average" },
  91	{ "rx_encaps_errors" },
  92	{ "tx_timeout_count" },
  93	{ "tx_restart_queue" },
  94	{ "tx_kicks" },
  95	{ "tx_flow_control_xon" },
  96	{ "tx_flow_control_xoff" },
  97	{ "rx_csum_offload_good" },
  98	{ "rx_csum_offload_errors"},
  99	{ "sg_ok"},
 100	{ "sg_linearized"},
 101};
 102
 103#define VECTOR_NUM_STATS	ARRAY_SIZE(ethtool_stats_keys)
 104
 105static void vector_reset_stats(struct vector_private *vp)
 106{
 107	/* We reuse the existing queue locks for stats */
 108
 109	/* RX stats are modified with RX head_lock held
 110	 * in vector_poll.
 111	 */
 112
 113	spin_lock(&vp->rx_queue->head_lock);
 114	vp->estats.rx_queue_max = 0;
 115	vp->estats.rx_queue_running_average = 0;
 116	vp->estats.rx_encaps_errors = 0;
 117	vp->estats.sg_ok = 0;
 118	vp->estats.sg_linearized = 0;
 119	spin_unlock(&vp->rx_queue->head_lock);
 120
 121	/* TX stats are modified with TX head_lock held
 122	 * in vector_send.
 123	 */
 124
 125	spin_lock(&vp->tx_queue->head_lock);
 126	vp->estats.tx_timeout_count = 0;
 127	vp->estats.tx_restart_queue = 0;
 128	vp->estats.tx_kicks = 0;
 129	vp->estats.tx_flow_control_xon = 0;
 130	vp->estats.tx_flow_control_xoff = 0;
 131	vp->estats.tx_queue_max = 0;
 132	vp->estats.tx_queue_running_average = 0;
 133	spin_unlock(&vp->tx_queue->head_lock);
 134}
 135
 136static int get_mtu(struct arglist *def)
 137{
 138	char *mtu = uml_vector_fetch_arg(def, "mtu");
 139	long result;
 140
 141	if (mtu != NULL) {
 142		if (kstrtoul(mtu, 10, &result) == 0)
 143			if ((result < (1 << 16) - 1) && (result >= 576))
 144				return result;
 145	}
 146	return ETH_MAX_PACKET;
 147}
 148
 149static char *get_bpf_file(struct arglist *def)
 150{
 151	return uml_vector_fetch_arg(def, "bpffile");
 152}
 153
 154static bool get_bpf_flash(struct arglist *def)
 155{
 156	char *allow = uml_vector_fetch_arg(def, "bpfflash");
 157	long result;
 158
 159	if (allow != NULL) {
 160		if (kstrtoul(allow, 10, &result) == 0)
 161			return result > 0;
 162	}
 163	return false;
 164}
 165
 166static int get_depth(struct arglist *def)
 167{
 168	char *mtu = uml_vector_fetch_arg(def, "depth");
 169	long result;
 170
 171	if (mtu != NULL) {
 172		if (kstrtoul(mtu, 10, &result) == 0)
 173			return result;
 174	}
 175	return DEFAULT_VECTOR_SIZE;
 176}
 177
 178static int get_headroom(struct arglist *def)
 179{
 180	char *mtu = uml_vector_fetch_arg(def, "headroom");
 181	long result;
 182
 183	if (mtu != NULL) {
 184		if (kstrtoul(mtu, 10, &result) == 0)
 185			return result;
 186	}
 187	return DEFAULT_HEADROOM;
 188}
 189
 190static int get_req_size(struct arglist *def)
 191{
 192	char *gro = uml_vector_fetch_arg(def, "gro");
 193	long result;
 194
 195	if (gro != NULL) {
 196		if (kstrtoul(gro, 10, &result) == 0) {
 197			if (result > 0)
 198				return 65536;
 199		}
 200	}
 201	return get_mtu(def) + ETH_HEADER_OTHER +
 202		get_headroom(def) + SAFETY_MARGIN;
 203}
 204
 205
 206static int get_transport_options(struct arglist *def)
 207{
 208	char *transport = uml_vector_fetch_arg(def, "transport");
 209	char *vector = uml_vector_fetch_arg(def, "vec");
 210
 211	int vec_rx = VECTOR_RX;
 212	int vec_tx = VECTOR_TX;
 213	long parsed;
 214	int result = 0;
 215
 216	if (transport == NULL)
 217		return -EINVAL;
 218
 219	if (vector != NULL) {
 220		if (kstrtoul(vector, 10, &parsed) == 0) {
 221			if (parsed == 0) {
 222				vec_rx = 0;
 223				vec_tx = 0;
 224			}
 225		}
 226	}
 227
 228	if (get_bpf_flash(def))
 229		result = VECTOR_BPF_FLASH;
 230
 231	if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
 232		return result;
 233	if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
 234		return (result | vec_rx | VECTOR_BPF);
 235	if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
 236		return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
 237	return (result | vec_rx | vec_tx);
 238}
 239
 240
 241/* A mini-buffer for packet drop read
 242 * All of our supported transports are datagram oriented and we always
 243 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
 244 * than the packet size it still counts as full packet read and will
 245 * clean the incoming stream to keep sigio/epoll happy
 246 */
 247
 248#define DROP_BUFFER_SIZE 32
 249
 250static char *drop_buffer;
 251
 252
 253/*
 254 * Advance the mmsg queue head by n = advance. Resets the queue to
 255 * maximum enqueue/dequeue-at-once capacity if possible. Called by
 256 * dequeuers. Caller must hold the head_lock!
 257 */
 258
 259static int vector_advancehead(struct vector_queue *qi, int advance)
 260{
 261	qi->head =
 262		(qi->head + advance)
 263			% qi->max_depth;
 264
 265
 266	atomic_sub(advance, &qi->queue_depth);
 267	return atomic_read(&qi->queue_depth);
 268}
 269
 270/*	Advance the queue tail by n = advance.
 271 *	This is called by enqueuers which should hold the
 272 *	head lock already
 273 */
 274
 275static int vector_advancetail(struct vector_queue *qi, int advance)
 276{
 277	qi->tail =
 278		(qi->tail + advance)
 279			% qi->max_depth;
 280	atomic_add(advance, &qi->queue_depth);
 281	return atomic_read(&qi->queue_depth);
 282}
 283
 284static int prep_msg(struct vector_private *vp,
 285	struct sk_buff *skb,
 286	struct iovec *iov)
 287{
 288	int iov_index = 0;
 289	int nr_frags, frag;
 290	skb_frag_t *skb_frag;
 291
 292	nr_frags = skb_shinfo(skb)->nr_frags;
 293	if (nr_frags > MAX_IOV_SIZE) {
 294		if (skb_linearize(skb) != 0)
 295			goto drop;
 296	}
 297	if (vp->header_size > 0) {
 298		iov[iov_index].iov_len = vp->header_size;
 299		vp->form_header(iov[iov_index].iov_base, skb, vp);
 300		iov_index++;
 301	}
 302	iov[iov_index].iov_base = skb->data;
 303	if (nr_frags > 0) {
 304		iov[iov_index].iov_len = skb->len - skb->data_len;
 305		vp->estats.sg_ok++;
 306	} else
 307		iov[iov_index].iov_len = skb->len;
 308	iov_index++;
 309	for (frag = 0; frag < nr_frags; frag++) {
 310		skb_frag = &skb_shinfo(skb)->frags[frag];
 311		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
 312		iov[iov_index].iov_len = skb_frag_size(skb_frag);
 313		iov_index++;
 314	}
 315	return iov_index;
 316drop:
 317	return -1;
 318}
 319/*
 320 * Generic vector enqueue with support for forming headers using transport
 321 * specific callback. Allows GRE, L2TPv3, RAW and other transports
 322 * to use a common enqueue procedure in vector mode
 323 */
 324
 325static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
 326{
 327	struct vector_private *vp = netdev_priv(qi->dev);
 328	int queue_depth;
 329	int packet_len;
 330	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
 331	int iov_count;
 332
 333	spin_lock(&qi->tail_lock);
 334	queue_depth = atomic_read(&qi->queue_depth);
 335
 336	if (skb)
 337		packet_len = skb->len;
 338
 339	if (queue_depth < qi->max_depth) {
 340
 341		*(qi->skbuff_vector + qi->tail) = skb;
 342		mmsg_vector += qi->tail;
 343		iov_count = prep_msg(
 344			vp,
 345			skb,
 346			mmsg_vector->msg_hdr.msg_iov
 347		);
 348		if (iov_count < 1)
 349			goto drop;
 350		mmsg_vector->msg_hdr.msg_iovlen = iov_count;
 351		mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
 352		mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
 353		wmb(); /* Make the packet visible to the NAPI poll thread */
 354		queue_depth = vector_advancetail(qi, 1);
 355	} else
 356		goto drop;
 357	spin_unlock(&qi->tail_lock);
 358	return queue_depth;
 359drop:
 360	qi->dev->stats.tx_dropped++;
 361	if (skb != NULL) {
 362		packet_len = skb->len;
 363		dev_consume_skb_any(skb);
 364		netdev_completed_queue(qi->dev, 1, packet_len);
 365	}
 366	spin_unlock(&qi->tail_lock);
 367	return queue_depth;
 368}
 369
 370static int consume_vector_skbs(struct vector_queue *qi, int count)
 371{
 372	struct sk_buff *skb;
 373	int skb_index;
 374	int bytes_compl = 0;
 375
 376	for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
 377		skb = *(qi->skbuff_vector + skb_index);
 378		/* mark as empty to ensure correct destruction if
 379		 * needed
 380		 */
 381		bytes_compl += skb->len;
 382		*(qi->skbuff_vector + skb_index) = NULL;
 383		dev_consume_skb_any(skb);
 384	}
 385	qi->dev->stats.tx_bytes += bytes_compl;
 386	qi->dev->stats.tx_packets += count;
 387	netdev_completed_queue(qi->dev, count, bytes_compl);
 388	return vector_advancehead(qi, count);
 389}
 390
 391/*
 392 * Generic vector dequeue via sendmmsg with support for forming headers
 393 * using transport specific callback. Allows GRE, L2TPv3, RAW and
 394 * other transports to use a common dequeue procedure in vector mode
 395 */
 396
 397
 398static int vector_send(struct vector_queue *qi)
 399{
 400	struct vector_private *vp = netdev_priv(qi->dev);
 401	struct mmsghdr *send_from;
 402	int result = 0, send_len;
 403
 404	if (spin_trylock(&qi->head_lock)) {
 405		/* update queue_depth to current value */
 406		while (atomic_read(&qi->queue_depth) > 0) {
 407			/* Calculate the start of the vector */
 408			send_len = atomic_read(&qi->queue_depth);
 409			send_from = qi->mmsg_vector;
 410			send_from += qi->head;
 411			/* Adjust vector size if wraparound */
 412			if (send_len + qi->head > qi->max_depth)
 413				send_len = qi->max_depth - qi->head;
 414			/* Try to TX as many packets as possible */
 415			if (send_len > 0) {
 416				result = uml_vector_sendmmsg(
 417					 vp->fds->tx_fd,
 418					 send_from,
 419					 send_len,
 420					 0
 421				);
 422				vp->in_write_poll =
 423					(result != send_len);
 424			}
 425			/* For some of the sendmmsg error scenarios
 426			 * we may end being unsure in the TX success
 427			 * for all packets. It is safer to declare
 428			 * them all TX-ed and blame the network.
 429			 */
 430			if (result < 0) {
 431				if (net_ratelimit())
 432					netdev_err(vp->dev, "sendmmsg err=%i\n",
 433						result);
 434				vp->in_error = true;
 435				result = send_len;
 436			}
 437			if (result > 0) {
 438				consume_vector_skbs(qi, result);
 439				/* This is equivalent to an TX IRQ.
 440				 * Restart the upper layers to feed us
 441				 * more packets.
 442				 */
 443				if (result > vp->estats.tx_queue_max)
 444					vp->estats.tx_queue_max = result;
 445				vp->estats.tx_queue_running_average =
 446					(vp->estats.tx_queue_running_average + result) >> 1;
 447			}
 448			netif_wake_queue(qi->dev);
 449			/* if TX is busy, break out of the send loop,
 450			 *  poll write IRQ will reschedule xmit for us.
 451			 */
 452			if (result != send_len) {
 453				vp->estats.tx_restart_queue++;
 454				break;
 455			}
 456		}
 457		spin_unlock(&qi->head_lock);
 458	}
 459	return atomic_read(&qi->queue_depth);
 460}
 461
 462/* Queue destructor. Deliberately stateless so we can use
 463 * it in queue cleanup if initialization fails.
 464 */
 465
 466static void destroy_queue(struct vector_queue *qi)
 467{
 468	int i;
 469	struct iovec *iov;
 470	struct vector_private *vp = netdev_priv(qi->dev);
 471	struct mmsghdr *mmsg_vector;
 472
 473	if (qi == NULL)
 474		return;
 475	/* deallocate any skbuffs - we rely on any unused to be
 476	 * set to NULL.
 477	 */
 478	if (qi->skbuff_vector != NULL) {
 479		for (i = 0; i < qi->max_depth; i++) {
 480			if (*(qi->skbuff_vector + i) != NULL)
 481				dev_kfree_skb_any(*(qi->skbuff_vector + i));
 482		}
 483		kfree(qi->skbuff_vector);
 484	}
 485	/* deallocate matching IOV structures including header buffs */
 486	if (qi->mmsg_vector != NULL) {
 487		mmsg_vector = qi->mmsg_vector;
 488		for (i = 0; i < qi->max_depth; i++) {
 489			iov = mmsg_vector->msg_hdr.msg_iov;
 490			if (iov != NULL) {
 491				if ((vp->header_size > 0) &&
 492					(iov->iov_base != NULL))
 493					kfree(iov->iov_base);
 494				kfree(iov);
 495			}
 496			mmsg_vector++;
 497		}
 498		kfree(qi->mmsg_vector);
 499	}
 500	kfree(qi);
 501}
 502
 503/*
 504 * Queue constructor. Create a queue with a given side.
 505 */
 506static struct vector_queue *create_queue(
 507	struct vector_private *vp,
 508	int max_size,
 509	int header_size,
 510	int num_extra_frags)
 511{
 512	struct vector_queue *result;
 513	int i;
 514	struct iovec *iov;
 515	struct mmsghdr *mmsg_vector;
 516
 517	result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
 518	if (result == NULL)
 519		return NULL;
 520	result->max_depth = max_size;
 521	result->dev = vp->dev;
 522	result->mmsg_vector = kmalloc(
 523		(sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
 524	if (result->mmsg_vector == NULL)
 525		goto out_mmsg_fail;
 526	result->skbuff_vector = kmalloc(
 527		(sizeof(void *) * max_size), GFP_KERNEL);
 528	if (result->skbuff_vector == NULL)
 529		goto out_skb_fail;
 530
 531	/* further failures can be handled safely by destroy_queue*/
 532
 533	mmsg_vector = result->mmsg_vector;
 534	for (i = 0; i < max_size; i++) {
 535		/* Clear all pointers - we use non-NULL as marking on
 536		 * what to free on destruction
 537		 */
 538		*(result->skbuff_vector + i) = NULL;
 539		mmsg_vector->msg_hdr.msg_iov = NULL;
 540		mmsg_vector++;
 541	}
 542	mmsg_vector = result->mmsg_vector;
 543	result->max_iov_frags = num_extra_frags;
 544	for (i = 0; i < max_size; i++) {
 545		if (vp->header_size > 0)
 546			iov = kmalloc_array(3 + num_extra_frags,
 547					    sizeof(struct iovec),
 548					    GFP_KERNEL
 549			);
 550		else
 551			iov = kmalloc_array(2 + num_extra_frags,
 552					    sizeof(struct iovec),
 553					    GFP_KERNEL
 554			);
 555		if (iov == NULL)
 556			goto out_fail;
 557		mmsg_vector->msg_hdr.msg_iov = iov;
 558		mmsg_vector->msg_hdr.msg_iovlen = 1;
 559		mmsg_vector->msg_hdr.msg_control = NULL;
 560		mmsg_vector->msg_hdr.msg_controllen = 0;
 561		mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
 562		mmsg_vector->msg_hdr.msg_name = NULL;
 563		mmsg_vector->msg_hdr.msg_namelen = 0;
 564		if (vp->header_size > 0) {
 565			iov->iov_base = kmalloc(header_size, GFP_KERNEL);
 566			if (iov->iov_base == NULL)
 567				goto out_fail;
 568			iov->iov_len = header_size;
 569			mmsg_vector->msg_hdr.msg_iovlen = 2;
 570			iov++;
 571		}
 572		iov->iov_base = NULL;
 573		iov->iov_len = 0;
 574		mmsg_vector++;
 575	}
 576	spin_lock_init(&result->head_lock);
 577	spin_lock_init(&result->tail_lock);
 578	atomic_set(&result->queue_depth, 0);
 579	result->head = 0;
 580	result->tail = 0;
 581	return result;
 582out_skb_fail:
 583	kfree(result->mmsg_vector);
 584out_mmsg_fail:
 585	kfree(result);
 586	return NULL;
 587out_fail:
 588	destroy_queue(result);
 589	return NULL;
 590}
 591
 592/*
 593 * We do not use the RX queue as a proper wraparound queue for now
 594 * This is not necessary because the consumption via napi_gro_receive()
 595 * happens in-line. While we can try using the return code of
 596 * netif_rx() for flow control there are no drivers doing this today.
 597 * For this RX specific use we ignore the tail/head locks and
 598 * just read into a prepared queue filled with skbuffs.
 599 */
 600
 601static struct sk_buff *prep_skb(
 602	struct vector_private *vp,
 603	struct user_msghdr *msg)
 604{
 605	int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
 606	struct sk_buff *result;
 607	int iov_index = 0, len;
 608	struct iovec *iov = msg->msg_iov;
 609	int err, nr_frags, frag;
 610	skb_frag_t *skb_frag;
 611
 612	if (vp->req_size <= linear)
 613		len = linear;
 614	else
 615		len = vp->req_size;
 616	result = alloc_skb_with_frags(
 617		linear,
 618		len - vp->max_packet,
 619		3,
 620		&err,
 621		GFP_ATOMIC
 622	);
 623	if (vp->header_size > 0)
 624		iov_index++;
 625	if (result == NULL) {
 626		iov[iov_index].iov_base = NULL;
 627		iov[iov_index].iov_len = 0;
 628		goto done;
 629	}
 630	skb_reserve(result, vp->headroom);
 631	result->dev = vp->dev;
 632	skb_put(result, vp->max_packet);
 633	result->data_len = len - vp->max_packet;
 634	result->len += len - vp->max_packet;
 635	skb_reset_mac_header(result);
 636	result->ip_summed = CHECKSUM_NONE;
 637	iov[iov_index].iov_base = result->data;
 638	iov[iov_index].iov_len = vp->max_packet;
 639	iov_index++;
 640
 641	nr_frags = skb_shinfo(result)->nr_frags;
 642	for (frag = 0; frag < nr_frags; frag++) {
 643		skb_frag = &skb_shinfo(result)->frags[frag];
 644		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
 645		if (iov[iov_index].iov_base != NULL)
 646			iov[iov_index].iov_len = skb_frag_size(skb_frag);
 647		else
 648			iov[iov_index].iov_len = 0;
 649		iov_index++;
 650	}
 651done:
 652	msg->msg_iovlen = iov_index;
 653	return result;
 654}
 655
 656
 657/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs */
 658
 659static void prep_queue_for_rx(struct vector_queue *qi)
 660{
 661	struct vector_private *vp = netdev_priv(qi->dev);
 662	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
 663	void **skbuff_vector = qi->skbuff_vector;
 664	int i, queue_depth;
 665
 666	queue_depth = atomic_read(&qi->queue_depth);
 667
 668	if (queue_depth == 0)
 669		return;
 670
 671	/* RX is always emptied 100% during each cycle, so we do not
 672	 * have to do the tail wraparound math for it.
 673	 */
 674
 675	qi->head = qi->tail = 0;
 676
 677	for (i = 0; i < queue_depth; i++) {
 678		/* it is OK if allocation fails - recvmmsg with NULL data in
 679		 * iov argument still performs an RX, just drops the packet
 680		 * This allows us stop faffing around with a "drop buffer"
 681		 */
 682
 683		*skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
 684		skbuff_vector++;
 685		mmsg_vector++;
 686	}
 687	atomic_set(&qi->queue_depth, 0);
 688}
 689
 690static struct vector_device *find_device(int n)
 691{
 692	struct vector_device *device;
 693	struct list_head *ele;
 694
 695	spin_lock(&vector_devices_lock);
 696	list_for_each(ele, &vector_devices) {
 697		device = list_entry(ele, struct vector_device, list);
 698		if (device->unit == n)
 699			goto out;
 700	}
 701	device = NULL;
 702 out:
 703	spin_unlock(&vector_devices_lock);
 704	return device;
 705}
 706
 707static int vector_parse(char *str, int *index_out, char **str_out,
 708			char **error_out)
 709{
 710	int n, err;
 711	char *start = str;
 712
 713	while ((*str != ':') && (strlen(str) > 1))
 714		str++;
 715	if (*str != ':') {
 716		*error_out = "Expected ':' after device number";
 717		return -EINVAL;
 718	}
 719	*str = '\0';
 720
 721	err = kstrtouint(start, 0, &n);
 722	if (err < 0) {
 723		*error_out = "Bad device number";
 724		return err;
 725	}
 726
 727	str++;
 728	if (find_device(n)) {
 729		*error_out = "Device already configured";
 730		return -EINVAL;
 731	}
 732
 733	*index_out = n;
 734	*str_out = str;
 735	return 0;
 736}
 737
 738static int vector_config(char *str, char **error_out)
 739{
 740	int err, n;
 741	char *params;
 742	struct arglist *parsed;
 743
 744	err = vector_parse(str, &n, &params, error_out);
 745	if (err != 0)
 746		return err;
 747
 748	/* This string is broken up and the pieces used by the underlying
 749	 * driver. We should copy it to make sure things do not go wrong
 750	 * later.
 751	 */
 752
 753	params = kstrdup(params, GFP_KERNEL);
 754	if (params == NULL) {
 755		*error_out = "vector_config failed to strdup string";
 756		return -ENOMEM;
 757	}
 758
 759	parsed = uml_parse_vector_ifspec(params);
 760
 761	if (parsed == NULL) {
 762		*error_out = "vector_config failed to parse parameters";
 763		kfree(params);
 764		return -EINVAL;
 765	}
 766
 767	vector_eth_configure(n, parsed);
 768	return 0;
 769}
 770
 771static int vector_id(char **str, int *start_out, int *end_out)
 772{
 773	char *end;
 774	int n;
 775
 776	n = simple_strtoul(*str, &end, 0);
 777	if ((*end != '\0') || (end == *str))
 778		return -1;
 779
 780	*start_out = n;
 781	*end_out = n;
 782	*str = end;
 783	return n;
 784}
 785
 786static int vector_remove(int n, char **error_out)
 787{
 788	struct vector_device *vec_d;
 789	struct net_device *dev;
 790	struct vector_private *vp;
 791
 792	vec_d = find_device(n);
 793	if (vec_d == NULL)
 794		return -ENODEV;
 795	dev = vec_d->dev;
 796	vp = netdev_priv(dev);
 797	if (vp->fds != NULL)
 798		return -EBUSY;
 799	unregister_netdev(dev);
 800	platform_device_unregister(&vec_d->pdev);
 801	return 0;
 802}
 803
 804/*
 805 * There is no shared per-transport initialization code, so
 806 * we will just initialize each interface one by one and
 807 * add them to a list
 808 */
 809
 810static struct platform_driver uml_net_driver = {
 811	.driver = {
 812		.name = DRIVER_NAME,
 813	},
 814};
 815
 816
 817static void vector_device_release(struct device *dev)
 818{
 819	struct vector_device *device =
 820		container_of(dev, struct vector_device, pdev.dev);
 821	struct net_device *netdev = device->dev;
 822
 823	list_del(&device->list);
 824	kfree(device);
 825	free_netdev(netdev);
 826}
 827
 828/* Bog standard recv using recvmsg - not used normally unless the user
 829 * explicitly specifies not to use recvmmsg vector RX.
 830 */
 831
 832static int vector_legacy_rx(struct vector_private *vp)
 833{
 834	int pkt_len;
 835	struct user_msghdr hdr;
 836	struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
 837	int iovpos = 0;
 838	struct sk_buff *skb;
 839	int header_check;
 840
 841	hdr.msg_name = NULL;
 842	hdr.msg_namelen = 0;
 843	hdr.msg_iov = (struct iovec *) &iov;
 844	hdr.msg_control = NULL;
 845	hdr.msg_controllen = 0;
 846	hdr.msg_flags = 0;
 847
 848	if (vp->header_size > 0) {
 849		iov[0].iov_base = vp->header_rxbuffer;
 850		iov[0].iov_len = vp->header_size;
 851	}
 852
 853	skb = prep_skb(vp, &hdr);
 854
 855	if (skb == NULL) {
 856		/* Read a packet into drop_buffer and don't do
 857		 * anything with it.
 858		 */
 859		iov[iovpos].iov_base = drop_buffer;
 860		iov[iovpos].iov_len = DROP_BUFFER_SIZE;
 861		hdr.msg_iovlen = 1;
 862		vp->dev->stats.rx_dropped++;
 863	}
 864
 865	pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
 866	if (pkt_len < 0) {
 867		vp->in_error = true;
 868		return pkt_len;
 869	}
 870
 871	if (skb != NULL) {
 872		if (pkt_len > vp->header_size) {
 873			if (vp->header_size > 0) {
 874				header_check = vp->verify_header(
 875					vp->header_rxbuffer, skb, vp);
 876				if (header_check < 0) {
 877					dev_kfree_skb_irq(skb);
 878					vp->dev->stats.rx_dropped++;
 879					vp->estats.rx_encaps_errors++;
 880					return 0;
 881				}
 882				if (header_check > 0) {
 883					vp->estats.rx_csum_offload_good++;
 884					skb->ip_summed = CHECKSUM_UNNECESSARY;
 885				}
 886			}
 887			pskb_trim(skb, pkt_len - vp->rx_header_size);
 888			skb->protocol = eth_type_trans(skb, skb->dev);
 889			vp->dev->stats.rx_bytes += skb->len;
 890			vp->dev->stats.rx_packets++;
 891			napi_gro_receive(&vp->napi, skb);
 892		} else {
 893			dev_kfree_skb_irq(skb);
 894		}
 895	}
 896	return pkt_len;
 897}
 898
 899/*
 900 * Packet at a time TX which falls back to vector TX if the
 901 * underlying transport is busy.
 902 */
 903
 904
 905
 906static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
 907{
 908	struct iovec iov[3 + MAX_IOV_SIZE];
 909	int iov_count, pkt_len = 0;
 910
 911	iov[0].iov_base = vp->header_txbuffer;
 912	iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
 913
 914	if (iov_count < 1)
 915		goto drop;
 916
 917	pkt_len = uml_vector_writev(
 918		vp->fds->tx_fd,
 919		(struct iovec *) &iov,
 920		iov_count
 921	);
 922
 923	if (pkt_len < 0)
 924		goto drop;
 925
 926	netif_trans_update(vp->dev);
 927	netif_wake_queue(vp->dev);
 928
 929	if (pkt_len > 0) {
 930		vp->dev->stats.tx_bytes += skb->len;
 931		vp->dev->stats.tx_packets++;
 932	} else {
 933		vp->dev->stats.tx_dropped++;
 934	}
 935	consume_skb(skb);
 936	return pkt_len;
 937drop:
 938	vp->dev->stats.tx_dropped++;
 939	consume_skb(skb);
 940	if (pkt_len < 0)
 941		vp->in_error = true;
 942	return pkt_len;
 943}
 944
 945/*
 946 * Receive as many messages as we can in one call using the special
 947 * mmsg vector matched to an skb vector which we prepared earlier.
 948 */
 949
 950static int vector_mmsg_rx(struct vector_private *vp, int budget)
 951{
 952	int packet_count, i;
 953	struct vector_queue *qi = vp->rx_queue;
 954	struct sk_buff *skb;
 955	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
 956	void **skbuff_vector = qi->skbuff_vector;
 957	int header_check;
 958
 959	/* Refresh the vector and make sure it is with new skbs and the
 960	 * iovs are updated to point to them.
 961	 */
 962
 963	prep_queue_for_rx(qi);
 964
 965	/* Fire the Lazy Gun - get as many packets as we can in one go. */
 966
 967	if (budget > qi->max_depth)
 968		budget = qi->max_depth;
 969
 970	packet_count = uml_vector_recvmmsg(
 971		vp->fds->rx_fd, qi->mmsg_vector, budget, 0);
 972
 973	if (packet_count < 0)
 974		vp->in_error = true;
 975
 976	if (packet_count <= 0)
 977		return packet_count;
 978
 979	/* We treat packet processing as enqueue, buffer refresh as dequeue
 980	 * The queue_depth tells us how many buffers have been used and how
 981	 * many do we need to prep the next time prep_queue_for_rx() is called.
 982	 */
 983
 984	atomic_add(packet_count, &qi->queue_depth);
 985
 986	for (i = 0; i < packet_count; i++) {
 987		skb = (*skbuff_vector);
 988		if (mmsg_vector->msg_len > vp->header_size) {
 989			if (vp->header_size > 0) {
 990				header_check = vp->verify_header(
 991					mmsg_vector->msg_hdr.msg_iov->iov_base,
 992					skb,
 993					vp
 994				);
 995				if (header_check < 0) {
 996				/* Overlay header failed to verify - discard.
 997				 * We can actually keep this skb and reuse it,
 998				 * but that will make the prep logic too
 999				 * complex.
1000				 */
1001					dev_kfree_skb_irq(skb);
1002					vp->estats.rx_encaps_errors++;
1003					continue;
1004				}
1005				if (header_check > 0) {
1006					vp->estats.rx_csum_offload_good++;
1007					skb->ip_summed = CHECKSUM_UNNECESSARY;
1008				}
1009			}
1010			pskb_trim(skb,
1011				mmsg_vector->msg_len - vp->rx_header_size);
1012			skb->protocol = eth_type_trans(skb, skb->dev);
1013			/*
1014			 * We do not need to lock on updating stats here
1015			 * The interrupt loop is non-reentrant.
1016			 */
1017			vp->dev->stats.rx_bytes += skb->len;
1018			vp->dev->stats.rx_packets++;
1019			napi_gro_receive(&vp->napi, skb);
1020		} else {
1021			/* Overlay header too short to do anything - discard.
1022			 * We can actually keep this skb and reuse it,
1023			 * but that will make the prep logic too complex.
1024			 */
1025			if (skb != NULL)
1026				dev_kfree_skb_irq(skb);
1027		}
1028		(*skbuff_vector) = NULL;
1029		/* Move to the next buffer element */
1030		mmsg_vector++;
1031		skbuff_vector++;
1032	}
1033	if (packet_count > 0) {
1034		if (vp->estats.rx_queue_max < packet_count)
1035			vp->estats.rx_queue_max = packet_count;
1036		vp->estats.rx_queue_running_average =
1037			(vp->estats.rx_queue_running_average + packet_count) >> 1;
1038	}
1039	return packet_count;
1040}
1041
1042static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1043{
1044	struct vector_private *vp = netdev_priv(dev);
1045	int queue_depth = 0;
1046
1047	if (vp->in_error) {
1048		deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1049		if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1050			deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1051		return NETDEV_TX_BUSY;
1052	}
1053
1054	if ((vp->options & VECTOR_TX) == 0) {
1055		writev_tx(vp, skb);
1056		return NETDEV_TX_OK;
1057	}
1058
1059	/* We do BQL only in the vector path, no point doing it in
1060	 * packet at a time mode as there is no device queue
1061	 */
1062
1063	netdev_sent_queue(vp->dev, skb->len);
1064	queue_depth = vector_enqueue(vp->tx_queue, skb);
1065
1066	if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1067		mod_timer(&vp->tl, vp->coalesce);
1068		return NETDEV_TX_OK;
1069	} else {
1070		queue_depth = vector_send(vp->tx_queue);
1071		if (queue_depth > 0)
1072			napi_schedule(&vp->napi);
1073	}
1074
1075	return NETDEV_TX_OK;
1076}
1077
1078static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1079{
1080	struct net_device *dev = dev_id;
1081	struct vector_private *vp = netdev_priv(dev);
1082
1083	if (!netif_running(dev))
1084		return IRQ_NONE;
1085	napi_schedule(&vp->napi);
1086	return IRQ_HANDLED;
1087
1088}
1089
1090static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1091{
1092	struct net_device *dev = dev_id;
1093	struct vector_private *vp = netdev_priv(dev);
1094
1095	if (!netif_running(dev))
1096		return IRQ_NONE;
1097	/* We need to pay attention to it only if we got
1098	 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1099	 * we ignore it. In the future, it may be worth
1100	 * it to improve the IRQ controller a bit to make
1101	 * tweaking the IRQ mask less costly
1102	 */
1103
1104	napi_schedule(&vp->napi);
1105	return IRQ_HANDLED;
1106
1107}
1108
1109static int irq_rr;
1110
1111static int vector_net_close(struct net_device *dev)
1112{
1113	struct vector_private *vp = netdev_priv(dev);
1114
1115	netif_stop_queue(dev);
1116	timer_delete(&vp->tl);
1117
1118	vp->opened = false;
1119
1120	if (vp->fds == NULL)
1121		return 0;
1122
1123	/* Disable and free all IRQS */
1124	if (vp->rx_irq > 0) {
1125		um_free_irq(vp->rx_irq, dev);
1126		vp->rx_irq = 0;
1127	}
1128	if (vp->tx_irq > 0) {
1129		um_free_irq(vp->tx_irq, dev);
1130		vp->tx_irq = 0;
1131	}
1132	napi_disable(&vp->napi);
1133	netif_napi_del(&vp->napi);
1134	if (vp->fds->rx_fd > 0) {
1135		if (vp->bpf)
1136			uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1137		os_close_file(vp->fds->rx_fd);
1138		vp->fds->rx_fd = -1;
1139	}
1140	if (vp->fds->tx_fd > 0) {
1141		os_close_file(vp->fds->tx_fd);
1142		vp->fds->tx_fd = -1;
1143	}
1144	if (vp->bpf != NULL)
1145		kfree(vp->bpf->filter);
1146	kfree(vp->bpf);
1147	vp->bpf = NULL;
1148	kfree(vp->fds->remote_addr);
1149	kfree(vp->transport_data);
1150	kfree(vp->header_rxbuffer);
1151	kfree(vp->header_txbuffer);
1152	if (vp->rx_queue != NULL)
1153		destroy_queue(vp->rx_queue);
1154	if (vp->tx_queue != NULL)
1155		destroy_queue(vp->tx_queue);
1156	kfree(vp->fds);
1157	vp->fds = NULL;
1158	vp->in_error = false;
1159	return 0;
1160}
1161
1162static int vector_poll(struct napi_struct *napi, int budget)
1163{
1164	struct vector_private *vp = container_of(napi, struct vector_private, napi);
1165	int work_done = 0;
1166	int err;
1167	bool tx_enqueued = false;
1168
1169	if ((vp->options & VECTOR_TX) != 0)
1170		tx_enqueued = (vector_send(vp->tx_queue) > 0);
1171	spin_lock(&vp->rx_queue->head_lock);
1172	if ((vp->options & VECTOR_RX) > 0)
1173		err = vector_mmsg_rx(vp, budget);
1174	else {
1175		err = vector_legacy_rx(vp);
1176		if (err > 0)
1177			err = 1;
1178	}
1179	spin_unlock(&vp->rx_queue->head_lock);
1180	if (err > 0)
1181		work_done += err;
1182
1183	if (tx_enqueued || err > 0)
1184		napi_schedule(napi);
1185	if (work_done <= budget)
1186		napi_complete_done(napi, work_done);
1187	return work_done;
1188}
1189
1190static void vector_reset_tx(struct work_struct *work)
1191{
1192	struct vector_private *vp =
1193		container_of(work, struct vector_private, reset_tx);
1194	netdev_reset_queue(vp->dev);
1195	netif_start_queue(vp->dev);
1196	netif_wake_queue(vp->dev);
1197}
1198
1199static int vector_net_open(struct net_device *dev)
1200{
1201	struct vector_private *vp = netdev_priv(dev);
1202	int err = -EINVAL;
1203	struct vector_device *vdevice;
1204
1205	if (vp->opened)
1206		return -ENXIO;
1207	vp->opened = true;
1208
1209	vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1210
1211	vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1212
1213	if (vp->fds == NULL)
1214		goto out_close;
1215
1216	if (build_transport_data(vp) < 0)
1217		goto out_close;
1218
1219	if ((vp->options & VECTOR_RX) > 0) {
1220		vp->rx_queue = create_queue(
1221			vp,
1222			get_depth(vp->parsed),
1223			vp->rx_header_size,
1224			MAX_IOV_SIZE
1225		);
1226		atomic_set(&vp->rx_queue->queue_depth, get_depth(vp->parsed));
1227	} else {
1228		vp->header_rxbuffer = kmalloc(
1229			vp->rx_header_size,
1230			GFP_KERNEL
1231		);
1232		if (vp->header_rxbuffer == NULL)
1233			goto out_close;
1234	}
1235	if ((vp->options & VECTOR_TX) > 0) {
1236		vp->tx_queue = create_queue(
1237			vp,
1238			get_depth(vp->parsed),
1239			vp->header_size,
1240			MAX_IOV_SIZE
1241		);
1242	} else {
1243		vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1244		if (vp->header_txbuffer == NULL)
1245			goto out_close;
1246	}
1247
1248	netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1249			      get_depth(vp->parsed));
1250	napi_enable(&vp->napi);
1251
1252	/* READ IRQ */
1253	err = um_request_irq(
1254		irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1255			IRQ_READ, vector_rx_interrupt,
1256			IRQF_SHARED, dev->name, dev);
1257	if (err < 0) {
1258		netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1259		err = -ENETUNREACH;
1260		goto out_close;
1261	}
1262	vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1263	dev->irq = irq_rr + VECTOR_BASE_IRQ;
1264	irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1265
1266	/* WRITE IRQ - we need it only if we have vector TX */
1267	if ((vp->options & VECTOR_TX) > 0) {
1268		err = um_request_irq(
1269			irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1270				IRQ_WRITE, vector_tx_interrupt,
1271				IRQF_SHARED, dev->name, dev);
1272		if (err < 0) {
1273			netdev_err(dev,
1274				"vector_open: failed to get tx irq(%d)\n", err);
1275			err = -ENETUNREACH;
1276			goto out_close;
1277		}
1278		vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1279		irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1280	}
1281
1282	if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1283		if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1284			vp->options |= VECTOR_BPF;
1285	}
1286	if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1287		vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1288
1289	if (vp->bpf != NULL)
1290		uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1291
1292	netif_start_queue(dev);
1293	vector_reset_stats(vp);
1294
1295	/* clear buffer - it can happen that the host side of the interface
1296	 * is full when we get here. In this case, new data is never queued,
1297	 * SIGIOs never arrive, and the net never works.
1298	 */
1299
1300	napi_schedule(&vp->napi);
1301
1302	vdevice = find_device(vp->unit);
1303	vdevice->opened = 1;
1304
1305	if ((vp->options & VECTOR_TX) != 0)
1306		add_timer(&vp->tl);
1307	return 0;
1308out_close:
1309	vector_net_close(dev);
1310	return err;
1311}
1312
1313
1314static void vector_net_set_multicast_list(struct net_device *dev)
1315{
1316	/* TODO: - we can do some BPF games here */
1317	return;
1318}
1319
1320static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1321{
1322	struct vector_private *vp = netdev_priv(dev);
1323
1324	vp->estats.tx_timeout_count++;
1325	netif_trans_update(dev);
1326	schedule_work(&vp->reset_tx);
1327}
1328
1329static netdev_features_t vector_fix_features(struct net_device *dev,
1330	netdev_features_t features)
1331{
1332	features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1333	return features;
1334}
1335
1336static int vector_set_features(struct net_device *dev,
1337	netdev_features_t features)
1338{
1339	struct vector_private *vp = netdev_priv(dev);
1340	/* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1341	 * no way to negotiate it on raw sockets, so we can change
1342	 * only our side.
1343	 */
1344	if (features & NETIF_F_GRO)
1345		/* All new frame buffers will be GRO-sized */
1346		vp->req_size = 65536;
1347	else
1348		/* All new frame buffers will be normal sized */
1349		vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1350	return 0;
1351}
1352
1353#ifdef CONFIG_NET_POLL_CONTROLLER
1354static void vector_net_poll_controller(struct net_device *dev)
1355{
1356	disable_irq(dev->irq);
1357	vector_rx_interrupt(dev->irq, dev);
1358	enable_irq(dev->irq);
1359}
1360#endif
1361
1362static void vector_net_get_drvinfo(struct net_device *dev,
1363				struct ethtool_drvinfo *info)
1364{
1365	strscpy(info->driver, DRIVER_NAME);
1366}
1367
1368static int vector_net_load_bpf_flash(struct net_device *dev,
1369				struct ethtool_flash *efl)
1370{
1371	struct vector_private *vp = netdev_priv(dev);
1372	struct vector_device *vdevice;
1373	const struct firmware *fw;
1374	int result = 0;
1375
1376	if (!(vp->options & VECTOR_BPF_FLASH)) {
1377		netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1378		return -1;
1379	}
1380
1381	if (vp->bpf != NULL) {
1382		if (vp->opened)
1383			uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1384		kfree(vp->bpf->filter);
1385		vp->bpf->filter = NULL;
1386	} else {
1387		vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1388		if (vp->bpf == NULL) {
1389			netdev_err(dev, "failed to allocate memory for firmware\n");
1390			goto flash_fail;
1391		}
1392	}
1393
1394	vdevice = find_device(vp->unit);
1395
1396	if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1397		goto flash_fail;
1398
1399	vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1400	if (!vp->bpf->filter)
1401		goto free_buffer;
1402
1403	vp->bpf->len = fw->size / sizeof(struct sock_filter);
1404	release_firmware(fw);
1405
1406	if (vp->opened)
1407		result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1408
1409	return result;
1410
1411free_buffer:
1412	release_firmware(fw);
1413
1414flash_fail:
1415	if (vp->bpf != NULL)
1416		kfree(vp->bpf->filter);
1417	kfree(vp->bpf);
1418	vp->bpf = NULL;
1419	return -1;
1420}
1421
1422static void vector_get_ringparam(struct net_device *netdev,
1423				 struct ethtool_ringparam *ring,
1424				 struct kernel_ethtool_ringparam *kernel_ring,
1425				 struct netlink_ext_ack *extack)
1426{
1427	struct vector_private *vp = netdev_priv(netdev);
1428
1429	ring->rx_max_pending = vp->rx_queue->max_depth;
1430	ring->tx_max_pending = vp->tx_queue->max_depth;
1431	ring->rx_pending = vp->rx_queue->max_depth;
1432	ring->tx_pending = vp->tx_queue->max_depth;
1433}
1434
1435static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1436{
1437	switch (stringset) {
1438	case ETH_SS_TEST:
1439		*buf = '\0';
1440		break;
1441	case ETH_SS_STATS:
1442		memcpy(buf, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
1443		break;
1444	default:
1445		WARN_ON(1);
1446		break;
1447	}
1448}
1449
1450static int vector_get_sset_count(struct net_device *dev, int sset)
1451{
1452	switch (sset) {
1453	case ETH_SS_TEST:
1454		return 0;
1455	case ETH_SS_STATS:
1456		return VECTOR_NUM_STATS;
1457	default:
1458		return -EOPNOTSUPP;
1459	}
1460}
1461
1462static void vector_get_ethtool_stats(struct net_device *dev,
1463	struct ethtool_stats *estats,
1464	u64 *tmp_stats)
1465{
1466	struct vector_private *vp = netdev_priv(dev);
1467
1468	/* Stats are modified in the dequeue portions of
1469	 * rx/tx which are protected by the head locks
1470	 * grabbing these locks here ensures they are up
1471	 * to date.
1472	 */
1473
1474	spin_lock(&vp->tx_queue->head_lock);
1475	spin_lock(&vp->rx_queue->head_lock);
1476	memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1477	spin_unlock(&vp->rx_queue->head_lock);
1478	spin_unlock(&vp->tx_queue->head_lock);
1479}
1480
1481static int vector_get_coalesce(struct net_device *netdev,
1482			       struct ethtool_coalesce *ec,
1483			       struct kernel_ethtool_coalesce *kernel_coal,
1484			       struct netlink_ext_ack *extack)
1485{
1486	struct vector_private *vp = netdev_priv(netdev);
1487
1488	ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1489	return 0;
1490}
1491
1492static int vector_set_coalesce(struct net_device *netdev,
1493			       struct ethtool_coalesce *ec,
1494			       struct kernel_ethtool_coalesce *kernel_coal,
1495			       struct netlink_ext_ack *extack)
1496{
1497	struct vector_private *vp = netdev_priv(netdev);
1498
1499	vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1500	if (vp->coalesce == 0)
1501		vp->coalesce = 1;
1502	return 0;
1503}
1504
1505static const struct ethtool_ops vector_net_ethtool_ops = {
1506	.supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1507	.get_drvinfo	= vector_net_get_drvinfo,
1508	.get_link	= ethtool_op_get_link,
1509	.get_ts_info	= ethtool_op_get_ts_info,
1510	.get_ringparam	= vector_get_ringparam,
1511	.get_strings	= vector_get_strings,
1512	.get_sset_count	= vector_get_sset_count,
1513	.get_ethtool_stats = vector_get_ethtool_stats,
1514	.get_coalesce	= vector_get_coalesce,
1515	.set_coalesce	= vector_set_coalesce,
1516	.flash_device	= vector_net_load_bpf_flash,
1517};
1518
1519
1520static const struct net_device_ops vector_netdev_ops = {
1521	.ndo_open		= vector_net_open,
1522	.ndo_stop		= vector_net_close,
1523	.ndo_start_xmit		= vector_net_start_xmit,
1524	.ndo_set_rx_mode	= vector_net_set_multicast_list,
1525	.ndo_tx_timeout		= vector_net_tx_timeout,
1526	.ndo_set_mac_address	= eth_mac_addr,
1527	.ndo_validate_addr	= eth_validate_addr,
1528	.ndo_fix_features	= vector_fix_features,
1529	.ndo_set_features	= vector_set_features,
1530#ifdef CONFIG_NET_POLL_CONTROLLER
1531	.ndo_poll_controller = vector_net_poll_controller,
1532#endif
1533};
1534
1535static void vector_timer_expire(struct timer_list *t)
1536{
1537	struct vector_private *vp = timer_container_of(vp, t, tl);
1538
1539	vp->estats.tx_kicks++;
1540	napi_schedule(&vp->napi);
1541}
1542
1543static void vector_setup_etheraddr(struct net_device *dev, char *str)
1544{
1545	u8 addr[ETH_ALEN];
1546
1547	if (str == NULL)
1548		goto random;
1549
1550	if (!mac_pton(str, addr)) {
1551		netdev_err(dev,
1552			"Failed to parse '%s' as an ethernet address\n", str);
1553		goto random;
1554	}
1555	if (is_multicast_ether_addr(addr)) {
1556		netdev_err(dev,
1557			"Attempt to assign a multicast ethernet address to a device disallowed\n");
1558		goto random;
1559	}
1560	if (!is_valid_ether_addr(addr)) {
1561		netdev_err(dev,
1562			"Attempt to assign an invalid ethernet address to a device disallowed\n");
1563		goto random;
1564	}
1565	if (!is_local_ether_addr(addr)) {
1566		netdev_warn(dev, "Warning: Assigning a globally valid ethernet address to a device\n");
1567		netdev_warn(dev, "You should set the 2nd rightmost bit in the first byte of the MAC,\n");
1568		netdev_warn(dev, "i.e. %02x:%02x:%02x:%02x:%02x:%02x\n",
1569			addr[0] | 0x02, addr[1], addr[2], addr[3], addr[4], addr[5]);
1570	}
1571	eth_hw_addr_set(dev, addr);
1572	return;
1573
1574random:
1575	netdev_info(dev, "Choosing a random ethernet address\n");
1576	eth_hw_addr_random(dev);
1577}
1578
1579static void vector_eth_configure(
1580		int n,
1581		struct arglist *def
1582	)
1583{
1584	struct vector_device *device;
1585	struct net_device *dev;
1586	struct vector_private *vp;
1587	int err;
1588
1589	device = kzalloc(sizeof(*device), GFP_KERNEL);
1590	if (device == NULL) {
1591		pr_err("Failed to allocate struct vector_device for vec%d\n", n);
1592		return;
1593	}
1594	dev = alloc_etherdev(sizeof(struct vector_private));
1595	if (dev == NULL) {
1596		pr_err("Failed to allocate struct net_device for vec%d\n", n);
1597		goto out_free_device;
1598	}
1599
1600	dev->mtu = get_mtu(def);
1601
1602	INIT_LIST_HEAD(&device->list);
1603	device->unit = n;
1604
1605	/* If this name ends up conflicting with an existing registered
1606	 * netdevice, that is OK, register_netdev{,ice}() will notice this
1607	 * and fail.
1608	 */
1609	snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1610	vector_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1611	vp = netdev_priv(dev);
1612
1613	/* sysfs register */
1614	if (!driver_registered) {
1615		platform_driver_register(&uml_net_driver);
1616		driver_registered = 1;
1617	}
1618	device->pdev.id = n;
1619	device->pdev.name = DRIVER_NAME;
1620	device->pdev.dev.release = vector_device_release;
1621	dev_set_drvdata(&device->pdev.dev, device);
1622	if (platform_device_register(&device->pdev))
1623		goto out_free_netdev;
1624	SET_NETDEV_DEV(dev, &device->pdev.dev);
1625
1626	device->dev = dev;
1627
1628	INIT_LIST_HEAD(&vp->list);
1629	vp->dev		= dev;
1630	vp->unit	= n;
1631	vp->options	= get_transport_options(def);
1632	vp->parsed	= def;
1633	vp->max_packet	= get_mtu(def) + ETH_HEADER_OTHER;
1634	/*
1635	 * TODO - we need to calculate headroom so that ip header
1636	 * is 16 byte aligned all the time
1637	 */
1638	vp->headroom	= get_headroom(def);
1639	vp->coalesce	= 2;
1640	vp->req_size	= get_req_size(def);
1641
1642	dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1643	INIT_WORK(&vp->reset_tx, vector_reset_tx);
1644
1645	timer_setup(&vp->tl, vector_timer_expire, 0);
1646
1647	/* FIXME */
1648	dev->netdev_ops = &vector_netdev_ops;
1649	dev->ethtool_ops = &vector_net_ethtool_ops;
1650	dev->watchdog_timeo = (HZ >> 1);
1651	/* primary IRQ - fixme */
1652	dev->irq = 0; /* we will adjust this once opened */
1653
1654	rtnl_lock();
1655	err = register_netdevice(dev);
1656	rtnl_unlock();
1657	if (err)
1658		goto out_undo_user_init;
1659
1660	spin_lock(&vector_devices_lock);
1661	list_add(&device->list, &vector_devices);
1662	spin_unlock(&vector_devices_lock);
1663
1664	return;
1665
1666out_undo_user_init:
1667	return;
1668out_free_netdev:
1669	free_netdev(dev);
1670out_free_device:
1671	kfree(device);
1672}
1673
1674
1675
1676
1677/*
1678 * Invoked late in the init
1679 */
1680
1681static int __init vector_init(void)
1682{
1683	struct list_head *ele;
1684	struct vector_cmd_line_arg *def;
1685	struct arglist *parsed;
1686
1687	list_for_each(ele, &vec_cmd_line) {
1688		def = list_entry(ele, struct vector_cmd_line_arg, list);
1689		parsed = uml_parse_vector_ifspec(def->arguments);
1690		if (parsed != NULL)
1691			vector_eth_configure(def->unit, parsed);
1692	}
1693	return 0;
1694}
1695
1696
1697/* Invoked at initial argument parsing, only stores
1698 * arguments until a proper vector_init is called
1699 * later
1700 */
1701
1702static int __init vector_setup(char *str)
1703{
1704	char *error;
1705	int n, err;
1706	struct vector_cmd_line_arg *new;
1707
1708	err = vector_parse(str, &n, &str, &error);
1709	if (err) {
1710		pr_err("Couldn't parse '%s': %s\n", str, error);
1711		return 1;
1712	}
1713	new = memblock_alloc_or_panic(sizeof(*new), SMP_CACHE_BYTES);
1714	INIT_LIST_HEAD(&new->list);
1715	new->unit = n;
1716	new->arguments = str;
1717	list_add_tail(&new->list, &vec_cmd_line);
1718	return 1;
1719}
1720
1721__setup("vec", vector_setup);
1722__uml_help(vector_setup,
1723"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1724"    Configure a vector io network device.\n\n"
1725);
1726
1727late_initcall(vector_init);
1728
1729static struct mc_device vector_mc = {
1730	.list		= LIST_HEAD_INIT(vector_mc.list),
1731	.name		= "vec",
1732	.config		= vector_config,
1733	.get_config	= NULL,
1734	.id		= vector_id,
1735	.remove		= vector_remove,
1736};
1737
1738#ifdef CONFIG_INET
1739static int vector_inetaddr_event(
1740	struct notifier_block *this,
1741	unsigned long event,
1742	void *ptr)
1743{
1744	return NOTIFY_DONE;
1745}
1746
1747static struct notifier_block vector_inetaddr_notifier = {
1748	.notifier_call		= vector_inetaddr_event,
1749};
1750
1751static void inet_register(void)
1752{
1753	register_inetaddr_notifier(&vector_inetaddr_notifier);
1754}
1755#else
1756static inline void inet_register(void)
1757{
1758}
1759#endif
1760
1761static int vector_net_init(void)
1762{
1763	mconsole_register_dev(&vector_mc);
1764	inet_register();
1765	return 0;
1766}
1767
1768__initcall(vector_net_init);
1769
1770
1771