Linux Audio

Check our new training course

Loading...
Note: File does not exist in v6.13.7.
   1/*
   2 * Copyright 2011 Tilera Corporation. All Rights Reserved.
   3 *
   4 *   This program is free software; you can redistribute it and/or
   5 *   modify it under the terms of the GNU General Public License
   6 *   as published by the Free Software Foundation, version 2.
   7 *
   8 *   This program is distributed in the hope that it will be useful, but
   9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11 *   NON INFRINGEMENT.  See the GNU General Public License for
  12 *   more details.
  13 */
  14
  15#include <linux/module.h>
  16#include <linux/init.h>
  17#include <linux/moduleparam.h>
  18#include <linux/sched.h>
  19#include <linux/kernel.h>      /* printk() */
  20#include <linux/slab.h>        /* kmalloc() */
  21#include <linux/errno.h>       /* error codes */
  22#include <linux/types.h>       /* size_t */
  23#include <linux/interrupt.h>
  24#include <linux/in.h>
  25#include <linux/netdevice.h>   /* struct device, and other headers */
  26#include <linux/etherdevice.h> /* eth_type_trans */
  27#include <linux/skbuff.h>
  28#include <linux/ioctl.h>
  29#include <linux/cdev.h>
  30#include <linux/hugetlb.h>
  31#include <linux/in6.h>
  32#include <linux/timer.h>
  33#include <linux/io.h>
  34#include <linux/u64_stats_sync.h>
  35#include <asm/checksum.h>
  36#include <asm/homecache.h>
  37
  38#include <hv/drv_xgbe_intf.h>
  39#include <hv/drv_xgbe_impl.h>
  40#include <hv/hypervisor.h>
  41#include <hv/netio_intf.h>
  42
  43/* For TSO */
  44#include <linux/ip.h>
  45#include <linux/tcp.h>
  46
  47
  48/*
  49 * First, "tile_net_init_module()" initializes all four "devices" which
  50 * can be used by linux.
  51 *
  52 * Then, "ifconfig DEVICE up" calls "tile_net_open()", which analyzes
  53 * the network cpus, then uses "tile_net_open_aux()" to initialize
  54 * LIPP/LEPP, and then uses "tile_net_open_inner()" to register all
  55 * the tiles, provide buffers to LIPP, allow ingress to start, and
  56 * turn on hypervisor interrupt handling (and NAPI) on all tiles.
  57 *
  58 * If registration fails due to the link being down, then "retry_work"
  59 * is used to keep calling "tile_net_open_inner()" until it succeeds.
  60 *
  61 * If "ifconfig DEVICE down" is called, it uses "tile_net_stop()" to
  62 * stop egress, drain the LIPP buffers, unregister all the tiles, stop
  63 * LIPP/LEPP, and wipe the LEPP queue.
  64 *
  65 * We start out with the ingress interrupt enabled on each CPU.  When
  66 * this interrupt fires, we disable it, and call "napi_schedule()".
  67 * This will cause "tile_net_poll()" to be called, which will pull
  68 * packets from the netio queue, filtering them out, or passing them
  69 * to "netif_receive_skb()".  If our budget is exhausted, we will
  70 * return, knowing we will be called again later.  Otherwise, we
  71 * reenable the ingress interrupt, and call "napi_complete()".
  72 *
  73 * HACK: Since disabling the ingress interrupt is not reliable, we
  74 * ignore the interrupt if the global "active" flag is false.
  75 *
  76 *
  77 * NOTE: The use of "native_driver" ensures that EPP exists, and that
  78 * we are using "LIPP" and "LEPP".
  79 *
  80 * NOTE: Failing to free completions for an arbitrarily long time
  81 * (which is defined to be illegal) does in fact cause bizarre
  82 * problems.  The "egress_timer" helps prevent this from happening.
  83 */
  84
  85
  86/* HACK: Allow use of "jumbo" packets. */
  87/* This should be 1500 if "jumbo" is not set in LIPP. */
  88/* This should be at most 10226 (10240 - 14) if "jumbo" is set in LIPP. */
  89/* ISSUE: This has not been thoroughly tested (except at 1500). */
  90#define TILE_NET_MTU 1500
  91
  92/* HACK: Define this to verify incoming packets. */
  93/* #define TILE_NET_VERIFY_INGRESS */
  94
  95/* Use 3000 to enable the Linux Traffic Control (QoS) layer, else 0. */
  96#define TILE_NET_TX_QUEUE_LEN 0
  97
  98/* Define to dump packets (prints out the whole packet on tx and rx). */
  99/* #define TILE_NET_DUMP_PACKETS */
 100
 101/* Define to enable debug spew (all PDEBUG's are enabled). */
 102/* #define TILE_NET_DEBUG */
 103
 104
 105/* Define to activate paranoia checks. */
 106/* #define TILE_NET_PARANOIA */
 107
 108/* Default transmit lockup timeout period, in jiffies. */
 109#define TILE_NET_TIMEOUT (5 * HZ)
 110
 111/* Default retry interval for bringing up the NetIO interface, in jiffies. */
 112#define TILE_NET_RETRY_INTERVAL (5 * HZ)
 113
 114/* Number of ports (xgbe0, xgbe1, gbe0, gbe1). */
 115#define TILE_NET_DEVS 4
 116
 117
 118
 119/* Paranoia. */
 120#if NET_IP_ALIGN != LIPP_PACKET_PADDING
 121#error "NET_IP_ALIGN must match LIPP_PACKET_PADDING."
 122#endif
 123
 124
 125/* Debug print. */
 126#ifdef TILE_NET_DEBUG
 127#define PDEBUG(fmt, args...) net_printk(fmt, ## args)
 128#else
 129#define PDEBUG(fmt, args...)
 130#endif
 131
 132
 133MODULE_AUTHOR("Tilera");
 134MODULE_LICENSE("GPL");
 135
 136
 137/*
 138 * Queue of incoming packets for a specific cpu and device.
 139 *
 140 * Includes a pointer to the "system" data, and the actual "user" data.
 141 */
 142struct tile_netio_queue {
 143	netio_queue_impl_t *__system_part;
 144	netio_queue_user_impl_t __user_part;
 145
 146};
 147
 148
 149/*
 150 * Statistics counters for a specific cpu and device.
 151 */
 152struct tile_net_stats_t {
 153	struct u64_stats_sync syncp;
 154	u64 rx_packets;		/* total packets received	*/
 155	u64 tx_packets;		/* total packets transmitted	*/
 156	u64 rx_bytes;		/* total bytes received 	*/
 157	u64 tx_bytes;		/* total bytes transmitted	*/
 158	u64 rx_errors;		/* packets truncated or marked bad by hw */
 159	u64 rx_dropped;		/* packets not for us or intf not up */
 160};
 161
 162
 163/*
 164 * Info for a specific cpu and device.
 165 *
 166 * ISSUE: There is a "dev" pointer in "napi" as well.
 167 */
 168struct tile_net_cpu {
 169	/* The NAPI struct. */
 170	struct napi_struct napi;
 171	/* Packet queue. */
 172	struct tile_netio_queue queue;
 173	/* Statistics. */
 174	struct tile_net_stats_t stats;
 175	/* True iff NAPI is enabled. */
 176	bool napi_enabled;
 177	/* True if this tile has successfully registered with the IPP. */
 178	bool registered;
 179	/* True if the link was down last time we tried to register. */
 180	bool link_down;
 181	/* True if "egress_timer" is scheduled. */
 182	bool egress_timer_scheduled;
 183	/* Number of small sk_buffs which must still be provided. */
 184	unsigned int num_needed_small_buffers;
 185	/* Number of large sk_buffs which must still be provided. */
 186	unsigned int num_needed_large_buffers;
 187	/* A timer for handling egress completions. */
 188	struct timer_list egress_timer;
 189};
 190
 191
 192/*
 193 * Info for a specific device.
 194 */
 195struct tile_net_priv {
 196	/* Our network device. */
 197	struct net_device *dev;
 198	/* Pages making up the egress queue. */
 199	struct page *eq_pages;
 200	/* Address of the actual egress queue. */
 201	lepp_queue_t *eq;
 202	/* Protects "eq". */
 203	spinlock_t eq_lock;
 204	/* The hypervisor handle for this interface. */
 205	int hv_devhdl;
 206	/* The intr bit mask that IDs this device. */
 207	u32 intr_id;
 208	/* True iff "tile_net_open_aux()" has succeeded. */
 209	bool partly_opened;
 210	/* True iff the device is "active". */
 211	bool active;
 212	/* Effective network cpus. */
 213	struct cpumask network_cpus_map;
 214	/* Number of network cpus. */
 215	int network_cpus_count;
 216	/* Credits per network cpu. */
 217	int network_cpus_credits;
 218	/* For NetIO bringup retries. */
 219	struct delayed_work retry_work;
 220	/* Quick access to per cpu data. */
 221	struct tile_net_cpu *cpu[NR_CPUS];
 222};
 223
 224/* Log2 of the number of small pages needed for the egress queue. */
 225#define EQ_ORDER  get_order(sizeof(lepp_queue_t))
 226/* Size of the egress queue's pages. */
 227#define EQ_SIZE   (1 << (PAGE_SHIFT + EQ_ORDER))
 228
 229/*
 230 * The actual devices (xgbe0, xgbe1, gbe0, gbe1).
 231 */
 232static struct net_device *tile_net_devs[TILE_NET_DEVS];
 233
 234/*
 235 * The "tile_net_cpu" structures for each device.
 236 */
 237static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe0);
 238static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe1);
 239static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe0);
 240static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe1);
 241
 242
 243/*
 244 * True if "network_cpus" was specified.
 245 */
 246static bool network_cpus_used;
 247
 248/*
 249 * The actual cpus in "network_cpus".
 250 */
 251static struct cpumask network_cpus_map;
 252
 253
 254
 255#ifdef TILE_NET_DEBUG
 256/*
 257 * printk with extra stuff.
 258 *
 259 * We print the CPU we're running in brackets.
 260 */
 261static void net_printk(char *fmt, ...)
 262{
 263	int i;
 264	int len;
 265	va_list args;
 266	static char buf[256];
 267
 268	len = sprintf(buf, "tile_net[%2.2d]: ", smp_processor_id());
 269	va_start(args, fmt);
 270	i = vscnprintf(buf + len, sizeof(buf) - len - 1, fmt, args);
 271	va_end(args);
 272	buf[255] = '\0';
 273	pr_notice(buf);
 274}
 275#endif
 276
 277
 278#ifdef TILE_NET_DUMP_PACKETS
 279/*
 280 * Dump a packet.
 281 */
 282static void dump_packet(unsigned char *data, unsigned long length, char *s)
 283{
 284	int my_cpu = smp_processor_id();
 285
 286	unsigned long i;
 287	char buf[128];
 288
 289	static unsigned int count;
 290
 291	pr_info("dump_packet(data %p, length 0x%lx s %s count 0x%x)\n",
 292	       data, length, s, count++);
 293
 294	pr_info("\n");
 295
 296	for (i = 0; i < length; i++) {
 297		if ((i & 0xf) == 0)
 298			sprintf(buf, "[%02d] %8.8lx:", my_cpu, i);
 299		sprintf(buf + strlen(buf), " %2.2x", data[i]);
 300		if ((i & 0xf) == 0xf || i == length - 1) {
 301			strcat(buf, "\n");
 302			pr_info("%s", buf);
 303		}
 304	}
 305}
 306#endif
 307
 308
 309/*
 310 * Provide support for the __netio_fastio1() swint
 311 * (see <hv/drv_xgbe_intf.h> for how it is used).
 312 *
 313 * The fastio swint2 call may clobber all the caller-saved registers.
 314 * It rarely clobbers memory, but we allow for the possibility in
 315 * the signature just to be on the safe side.
 316 *
 317 * Also, gcc doesn't seem to allow an input operand to be
 318 * clobbered, so we fake it with dummy outputs.
 319 *
 320 * This function can't be static because of the way it is declared
 321 * in the netio header.
 322 */
 323inline int __netio_fastio1(u32 fastio_index, u32 arg0)
 324{
 325	long result, clobber_r1, clobber_r10;
 326	asm volatile("swint2"
 327		     : "=R00" (result),
 328		       "=R01" (clobber_r1), "=R10" (clobber_r10)
 329		     : "R10" (fastio_index), "R01" (arg0)
 330		     : "memory", "r2", "r3", "r4",
 331		       "r5", "r6", "r7", "r8", "r9",
 332		       "r11", "r12", "r13", "r14",
 333		       "r15", "r16", "r17", "r18", "r19",
 334		       "r20", "r21", "r22", "r23", "r24",
 335		       "r25", "r26", "r27", "r28", "r29");
 336	return result;
 337}
 338
 339
 340static void tile_net_return_credit(struct tile_net_cpu *info)
 341{
 342	struct tile_netio_queue *queue = &info->queue;
 343	netio_queue_user_impl_t *qup = &queue->__user_part;
 344
 345	/* Return four credits after every fourth packet. */
 346	if (--qup->__receive_credit_remaining == 0) {
 347		u32 interval = qup->__receive_credit_interval;
 348		qup->__receive_credit_remaining = interval;
 349		__netio_fastio_return_credits(qup->__fastio_index, interval);
 350	}
 351}
 352
 353
 354
 355/*
 356 * Provide a linux buffer to LIPP.
 357 */
 358static void tile_net_provide_linux_buffer(struct tile_net_cpu *info,
 359					  void *va, bool small)
 360{
 361	struct tile_netio_queue *queue = &info->queue;
 362
 363	/* Convert "va" and "small" to "linux_buffer_t". */
 364	unsigned int buffer = ((unsigned int)(__pa(va) >> 7) << 1) + small;
 365
 366	__netio_fastio_free_buffer(queue->__user_part.__fastio_index, buffer);
 367}
 368
 369
 370/*
 371 * Provide a linux buffer for LIPP.
 372 *
 373 * Note that the ACTUAL allocation for each buffer is a "struct sk_buff",
 374 * plus a chunk of memory that includes not only the requested bytes, but
 375 * also NET_SKB_PAD bytes of initial padding, and a "struct skb_shared_info".
 376 *
 377 * Note that "struct skb_shared_info" is 88 bytes with 64K pages and
 378 * 268 bytes with 4K pages (since the frags[] array needs 18 entries).
 379 *
 380 * Without jumbo packets, the maximum packet size will be 1536 bytes,
 381 * and we use 2 bytes (NET_IP_ALIGN) of padding.  ISSUE: If we told
 382 * the hardware to clip at 1518 bytes instead of 1536 bytes, then we
 383 * could save an entire cache line, but in practice, we don't need it.
 384 *
 385 * Since CPAs are 38 bits, and we can only encode the high 31 bits in
 386 * a "linux_buffer_t", the low 7 bits must be zero, and thus, we must
 387 * align the actual "va" mod 128.
 388 *
 389 * We assume that the underlying "head" will be aligned mod 64.  Note
 390 * that in practice, we have seen "head" NOT aligned mod 128 even when
 391 * using 2048 byte allocations, which is surprising.
 392 *
 393 * If "head" WAS always aligned mod 128, we could change LIPP to
 394 * assume that the low SIX bits are zero, and the 7th bit is one, that
 395 * is, align the actual "va" mod 128 plus 64, which would be "free".
 396 *
 397 * For now, the actual "head" pointer points at NET_SKB_PAD bytes of
 398 * padding, plus 28 or 92 bytes of extra padding, plus the sk_buff
 399 * pointer, plus the NET_IP_ALIGN padding, plus 126 or 1536 bytes for
 400 * the actual packet, plus 62 bytes of empty padding, plus some
 401 * padding and the "struct skb_shared_info".
 402 *
 403 * With 64K pages, a large buffer thus needs 32+92+4+2+1536+62+88
 404 * bytes, or 1816 bytes, which fits comfortably into 2048 bytes.
 405 *
 406 * With 64K pages, a small buffer thus needs 32+92+4+2+126+88
 407 * bytes, or 344 bytes, which means we are wasting 64+ bytes, and
 408 * could presumably increase the size of small buffers.
 409 *
 410 * With 4K pages, a large buffer thus needs 32+92+4+2+1536+62+268
 411 * bytes, or 1996 bytes, which fits comfortably into 2048 bytes.
 412 *
 413 * With 4K pages, a small buffer thus needs 32+92+4+2+126+268
 414 * bytes, or 524 bytes, which is annoyingly wasteful.
 415 *
 416 * Maybe we should increase LIPP_SMALL_PACKET_SIZE to 192?
 417 *
 418 * ISSUE: Maybe we should increase "NET_SKB_PAD" to 64?
 419 */
 420static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
 421					   bool small)
 422{
 423#if TILE_NET_MTU <= 1536
 424	/* Without "jumbo", 2 + 1536 should be sufficient. */
 425	unsigned int large_size = NET_IP_ALIGN + 1536;
 426#else
 427	/* ISSUE: This has not been tested. */
 428	unsigned int large_size = NET_IP_ALIGN + TILE_NET_MTU + 100;
 429#endif
 430
 431	/* Avoid "false sharing" with last cache line. */
 432	/* ISSUE: This is already done by "netdev_alloc_skb()". */
 433	unsigned int len =
 434		 (((small ? LIPP_SMALL_PACKET_SIZE : large_size) +
 435		   CHIP_L2_LINE_SIZE() - 1) & -CHIP_L2_LINE_SIZE());
 436
 437	unsigned int padding = 128 - NET_SKB_PAD;
 438	unsigned int align;
 439
 440	struct sk_buff *skb;
 441	void *va;
 442
 443	struct sk_buff **skb_ptr;
 444
 445	/* Request 96 extra bytes for alignment purposes. */
 446	skb = netdev_alloc_skb(info->napi.dev, len + padding);
 447	if (skb == NULL)
 448		return false;
 449
 450	/* Skip 32 or 96 bytes to align "data" mod 128. */
 451	align = -(long)skb->data & (128 - 1);
 452	BUG_ON(align > padding);
 453	skb_reserve(skb, align);
 454
 455	/* This address is given to IPP. */
 456	va = skb->data;
 457
 458	/* Buffers must not span a huge page. */
 459	BUG_ON(((((long)va & ~HPAGE_MASK) + len) & HPAGE_MASK) != 0);
 460
 461#ifdef TILE_NET_PARANOIA
 462#if CHIP_HAS_CBOX_HOME_MAP()
 463	if (hash_default) {
 464		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
 465		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
 466			panic("Non-HFH ingress buffer! VA=%p Mode=%d PTE=%llx",
 467			      va, hv_pte_get_mode(pte), hv_pte_val(pte));
 468	}
 469#endif
 470#endif
 471
 472	/* Invalidate the packet buffer. */
 473	if (!hash_default)
 474		__inv_buffer(va, len);
 475
 476	/* Skip two bytes to satisfy LIPP assumptions. */
 477	/* Note that this aligns IP on a 16 byte boundary. */
 478	/* ISSUE: Do this when the packet arrives? */
 479	skb_reserve(skb, NET_IP_ALIGN);
 480
 481	/* Save a back-pointer to 'skb'. */
 482	skb_ptr = va - sizeof(*skb_ptr);
 483	*skb_ptr = skb;
 484
 485	/* Make sure "skb_ptr" has been flushed. */
 486	__insn_mf();
 487
 488	/* Provide the new buffer. */
 489	tile_net_provide_linux_buffer(info, va, small);
 490
 491	return true;
 492}
 493
 494
 495/*
 496 * Provide linux buffers for LIPP.
 497 */
 498static void tile_net_provide_needed_buffers(struct tile_net_cpu *info)
 499{
 500	while (info->num_needed_small_buffers != 0) {
 501		if (!tile_net_provide_needed_buffer(info, true))
 502			goto oops;
 503		info->num_needed_small_buffers--;
 504	}
 505
 506	while (info->num_needed_large_buffers != 0) {
 507		if (!tile_net_provide_needed_buffer(info, false))
 508			goto oops;
 509		info->num_needed_large_buffers--;
 510	}
 511
 512	return;
 513
 514oops:
 515
 516	/* Add a description to the page allocation failure dump. */
 517	pr_notice("Could not provide a linux buffer to LIPP.\n");
 518}
 519
 520
 521/*
 522 * Grab some LEPP completions, and store them in "comps", of size
 523 * "comps_size", and return the number of completions which were
 524 * stored, so the caller can free them.
 525 */
 526static unsigned int tile_net_lepp_grab_comps(lepp_queue_t *eq,
 527					     struct sk_buff *comps[],
 528					     unsigned int comps_size,
 529					     unsigned int min_size)
 530{
 531	unsigned int n = 0;
 532
 533	unsigned int comp_head = eq->comp_head;
 534	unsigned int comp_busy = eq->comp_busy;
 535
 536	while (comp_head != comp_busy && n < comps_size) {
 537		comps[n++] = eq->comps[comp_head];
 538		LEPP_QINC(comp_head);
 539	}
 540
 541	if (n < min_size)
 542		return 0;
 543
 544	eq->comp_head = comp_head;
 545
 546	return n;
 547}
 548
 549
 550/*
 551 * Free some comps, and return true iff there are still some pending.
 552 */
 553static bool tile_net_lepp_free_comps(struct net_device *dev, bool all)
 554{
 555	struct tile_net_priv *priv = netdev_priv(dev);
 556
 557	lepp_queue_t *eq = priv->eq;
 558
 559	struct sk_buff *olds[64];
 560	unsigned int wanted = 64;
 561	unsigned int i, n;
 562	bool pending;
 563
 564	spin_lock(&priv->eq_lock);
 565
 566	if (all)
 567		eq->comp_busy = eq->comp_tail;
 568
 569	n = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
 570
 571	pending = (eq->comp_head != eq->comp_tail);
 572
 573	spin_unlock(&priv->eq_lock);
 574
 575	for (i = 0; i < n; i++)
 576		kfree_skb(olds[i]);
 577
 578	return pending;
 579}
 580
 581
 582/*
 583 * Make sure the egress timer is scheduled.
 584 *
 585 * Note that we use "schedule if not scheduled" logic instead of the more
 586 * obvious "reschedule" logic, because "reschedule" is fairly expensive.
 587 */
 588static void tile_net_schedule_egress_timer(struct tile_net_cpu *info)
 589{
 590	if (!info->egress_timer_scheduled) {
 591		mod_timer_pinned(&info->egress_timer, jiffies + 1);
 592		info->egress_timer_scheduled = true;
 593	}
 594}
 595
 596
 597/*
 598 * The "function" for "info->egress_timer".
 599 *
 600 * This timer will reschedule itself as long as there are any pending
 601 * completions expected (on behalf of any tile).
 602 *
 603 * ISSUE: Realistically, will the timer ever stop scheduling itself?
 604 *
 605 * ISSUE: This timer is almost never actually needed, so just use a global
 606 * timer that can run on any tile.
 607 *
 608 * ISSUE: Maybe instead track number of expected completions, and free
 609 * only that many, resetting to zero if "pending" is ever false.
 610 */
 611static void tile_net_handle_egress_timer(unsigned long arg)
 612{
 613	struct tile_net_cpu *info = (struct tile_net_cpu *)arg;
 614	struct net_device *dev = info->napi.dev;
 615
 616	/* The timer is no longer scheduled. */
 617	info->egress_timer_scheduled = false;
 618
 619	/* Free comps, and reschedule timer if more are pending. */
 620	if (tile_net_lepp_free_comps(dev, false))
 621		tile_net_schedule_egress_timer(info);
 622}
 623
 624
 625static void tile_net_discard_aux(struct tile_net_cpu *info, int index)
 626{
 627	struct tile_netio_queue *queue = &info->queue;
 628	netio_queue_impl_t *qsp = queue->__system_part;
 629	netio_queue_user_impl_t *qup = &queue->__user_part;
 630
 631	int index2_aux = index + sizeof(netio_pkt_t);
 632	int index2 =
 633		((index2_aux ==
 634		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
 635		 0 : index2_aux);
 636
 637	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
 638
 639	/* Extract the "linux_buffer_t". */
 640	unsigned int buffer = pkt->__packet.word;
 641
 642	/* Convert "linux_buffer_t" to "va". */
 643	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
 644
 645	/* Acquire the associated "skb". */
 646	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
 647	struct sk_buff *skb = *skb_ptr;
 648
 649	kfree_skb(skb);
 650
 651	/* Consume this packet. */
 652	qup->__packet_receive_read = index2;
 653}
 654
 655
 656/*
 657 * Like "tile_net_poll()", but just discard packets.
 658 */
 659static void tile_net_discard_packets(struct net_device *dev)
 660{
 661	struct tile_net_priv *priv = netdev_priv(dev);
 662	int my_cpu = smp_processor_id();
 663	struct tile_net_cpu *info = priv->cpu[my_cpu];
 664	struct tile_netio_queue *queue = &info->queue;
 665	netio_queue_impl_t *qsp = queue->__system_part;
 666	netio_queue_user_impl_t *qup = &queue->__user_part;
 667
 668	while (qup->__packet_receive_read !=
 669	       qsp->__packet_receive_queue.__packet_write) {
 670		int index = qup->__packet_receive_read;
 671		tile_net_discard_aux(info, index);
 672	}
 673}
 674
 675
 676/*
 677 * Handle the next packet.  Return true if "processed", false if "filtered".
 678 */
 679static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 680{
 681	struct net_device *dev = info->napi.dev;
 682
 683	struct tile_netio_queue *queue = &info->queue;
 684	netio_queue_impl_t *qsp = queue->__system_part;
 685	netio_queue_user_impl_t *qup = &queue->__user_part;
 686	struct tile_net_stats_t *stats = &info->stats;
 687
 688	int filter;
 689
 690	int index2_aux = index + sizeof(netio_pkt_t);
 691	int index2 =
 692		((index2_aux ==
 693		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
 694		 0 : index2_aux);
 695
 696	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
 697
 698	netio_pkt_metadata_t *metadata = NETIO_PKT_METADATA(pkt);
 699	netio_pkt_status_t pkt_status = NETIO_PKT_STATUS_M(metadata, pkt);
 700
 701	/* Extract the packet size.  FIXME: Shouldn't the second line */
 702	/* get subtracted?  Mostly moot, since it should be "zero". */
 703	unsigned long len =
 704		(NETIO_PKT_CUSTOM_LENGTH(pkt) +
 705		 NET_IP_ALIGN - NETIO_PACKET_PADDING);
 706
 707	/* Extract the "linux_buffer_t". */
 708	unsigned int buffer = pkt->__packet.word;
 709
 710	/* Extract "small" (vs "large"). */
 711	bool small = ((buffer & 1) != 0);
 712
 713	/* Convert "linux_buffer_t" to "va". */
 714	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
 715
 716	/* Extract the packet data pointer. */
 717	/* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
 718	unsigned char *buf = va + NET_IP_ALIGN;
 719
 720	/* Invalidate the packet buffer. */
 721	if (!hash_default)
 722		__inv_buffer(buf, len);
 723
 724	/* ISSUE: Is this needed? */
 725	dev->last_rx = jiffies;
 726
 727#ifdef TILE_NET_DUMP_PACKETS
 728	dump_packet(buf, len, "rx");
 729#endif /* TILE_NET_DUMP_PACKETS */
 730
 731#ifdef TILE_NET_VERIFY_INGRESS
 732	if (pkt_status == NETIO_PKT_STATUS_OVERSIZE && len >= 64) {
 733		dump_packet(buf, len, "rx");
 734		panic("Unexpected OVERSIZE.");
 735	}
 736#endif
 737
 738	filter = 0;
 739
 740	if (pkt_status == NETIO_PKT_STATUS_BAD) {
 741		/* Handle CRC error and hardware truncation. */
 742		filter = 2;
 743	} else if (!(dev->flags & IFF_UP)) {
 744		/* Filter packets received before we're up. */
 745		filter = 1;
 746	} else if (NETIO_PKT_ETHERTYPE_RECOGNIZED_M(metadata, pkt) &&
 747		   pkt_status == NETIO_PKT_STATUS_UNDERSIZE) {
 748		/* Filter "truncated" packets. */
 749		filter = 2;
 750	} else if (!(dev->flags & IFF_PROMISC)) {
 751		if (!is_multicast_ether_addr(buf)) {
 752			/* Filter packets not for our address. */
 753			const u8 *mine = dev->dev_addr;
 754			filter = !ether_addr_equal(mine, buf);
 755		}
 756	}
 757
 758	u64_stats_update_begin(&stats->syncp);
 759
 760	if (filter != 0) {
 761
 762		if (filter == 1)
 763			stats->rx_dropped++;
 764		else
 765			stats->rx_errors++;
 766
 767		tile_net_provide_linux_buffer(info, va, small);
 768
 769	} else {
 770
 771		/* Acquire the associated "skb". */
 772		struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
 773		struct sk_buff *skb = *skb_ptr;
 774
 775		/* Paranoia. */
 776		if (skb->data != buf)
 777			panic("Corrupt linux buffer from LIPP! "
 778			      "VA=%p, skb=%p, skb->data=%p\n",
 779			      va, skb, skb->data);
 780
 781		/* Encode the actual packet length. */
 782		skb_put(skb, len);
 783
 784		/* NOTE: This call also sets "skb->dev = dev". */
 785		skb->protocol = eth_type_trans(skb, dev);
 786
 787		/* Avoid recomputing "good" TCP/UDP checksums. */
 788		if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
 789			skb->ip_summed = CHECKSUM_UNNECESSARY;
 790
 791		netif_receive_skb(skb);
 792
 793		stats->rx_packets++;
 794		stats->rx_bytes += len;
 795	}
 796
 797	u64_stats_update_end(&stats->syncp);
 798
 799	/* ISSUE: It would be nice to defer this until the packet has */
 800	/* actually been processed. */
 801	tile_net_return_credit(info);
 802
 803	/* Consume this packet. */
 804	qup->__packet_receive_read = index2;
 805
 806	return !filter;
 807}
 808
 809
 810/*
 811 * Handle some packets for the given device on the current CPU.
 812 *
 813 * If "tile_net_stop()" is called on some other tile while this
 814 * function is running, we will return, hopefully before that
 815 * other tile asks us to call "napi_disable()".
 816 *
 817 * The "rotting packet" race condition occurs if a packet arrives
 818 * during the extremely narrow window between the queue appearing to
 819 * be empty, and the ingress interrupt being re-enabled.  This happens
 820 * a LOT under heavy network load.
 821 */
 822static int tile_net_poll(struct napi_struct *napi, int budget)
 823{
 824	struct net_device *dev = napi->dev;
 825	struct tile_net_priv *priv = netdev_priv(dev);
 826	int my_cpu = smp_processor_id();
 827	struct tile_net_cpu *info = priv->cpu[my_cpu];
 828	struct tile_netio_queue *queue = &info->queue;
 829	netio_queue_impl_t *qsp = queue->__system_part;
 830	netio_queue_user_impl_t *qup = &queue->__user_part;
 831
 832	unsigned int work = 0;
 833
 834	if (budget <= 0)
 835		goto done;
 836
 837	while (priv->active) {
 838		int index = qup->__packet_receive_read;
 839		if (index == qsp->__packet_receive_queue.__packet_write)
 840			break;
 841
 842		if (tile_net_poll_aux(info, index)) {
 843			if (++work >= budget)
 844				goto done;
 845		}
 846	}
 847
 848	napi_complete(&info->napi);
 849
 850	if (!priv->active)
 851		goto done;
 852
 853	/* Re-enable the ingress interrupt. */
 854	enable_percpu_irq(priv->intr_id, 0);
 855
 856	/* HACK: Avoid the "rotting packet" problem (see above). */
 857	if (qup->__packet_receive_read !=
 858	    qsp->__packet_receive_queue.__packet_write) {
 859		/* ISSUE: Sometimes this returns zero, presumably */
 860		/* because an interrupt was handled for this tile. */
 861		(void)napi_reschedule(&info->napi);
 862	}
 863
 864done:
 865
 866	if (priv->active)
 867		tile_net_provide_needed_buffers(info);
 868
 869	return work;
 870}
 871
 872
 873/*
 874 * Handle an ingress interrupt for the given device on the current cpu.
 875 *
 876 * ISSUE: Sometimes this gets called after "disable_percpu_irq()" has
 877 * been called!  This is probably due to "pending hypervisor downcalls".
 878 *
 879 * ISSUE: Is there any race condition between the "napi_schedule()" here
 880 * and the "napi_complete()" call above?
 881 */
 882static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
 883{
 884	struct net_device *dev = (struct net_device *)dev_ptr;
 885	struct tile_net_priv *priv = netdev_priv(dev);
 886	int my_cpu = smp_processor_id();
 887	struct tile_net_cpu *info = priv->cpu[my_cpu];
 888
 889	/* Disable the ingress interrupt. */
 890	disable_percpu_irq(priv->intr_id);
 891
 892	/* Ignore unwanted interrupts. */
 893	if (!priv->active)
 894		return IRQ_HANDLED;
 895
 896	/* ISSUE: Sometimes "info->napi_enabled" is false here. */
 897
 898	napi_schedule(&info->napi);
 899
 900	return IRQ_HANDLED;
 901}
 902
 903
 904/*
 905 * One time initialization per interface.
 906 */
 907static int tile_net_open_aux(struct net_device *dev)
 908{
 909	struct tile_net_priv *priv = netdev_priv(dev);
 910
 911	int ret;
 912	int dummy;
 913	unsigned int epp_lotar;
 914
 915	/*
 916	 * Find out where EPP memory should be homed.
 917	 */
 918	ret = hv_dev_pread(priv->hv_devhdl, 0,
 919			   (HV_VirtAddr)&epp_lotar, sizeof(epp_lotar),
 920			   NETIO_EPP_SHM_OFF);
 921	if (ret < 0) {
 922		pr_err("could not read epp_shm_queue lotar.\n");
 923		return -EIO;
 924	}
 925
 926	/*
 927	 * Home the page on the EPP.
 928	 */
 929	{
 930		int epp_home = hv_lotar_to_cpu(epp_lotar);
 931		homecache_change_page_home(priv->eq_pages, EQ_ORDER, epp_home);
 932	}
 933
 934	/*
 935	 * Register the EPP shared memory queue.
 936	 */
 937	{
 938		netio_ipp_address_t ea = {
 939			.va = 0,
 940			.pa = __pa(priv->eq),
 941			.pte = hv_pte(0),
 942			.size = EQ_SIZE,
 943		};
 944		ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
 945		ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
 946		ret = hv_dev_pwrite(priv->hv_devhdl, 0,
 947				    (HV_VirtAddr)&ea,
 948				    sizeof(ea),
 949				    NETIO_EPP_SHM_OFF);
 950		if (ret < 0)
 951			return -EIO;
 952	}
 953
 954	/*
 955	 * Start LIPP/LEPP.
 956	 */
 957	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
 958			  sizeof(dummy), NETIO_IPP_START_SHIM_OFF) < 0) {
 959		pr_warning("Failed to start LIPP/LEPP.\n");
 960		return -EIO;
 961	}
 962
 963	return 0;
 964}
 965
 966
 967/*
 968 * Register with hypervisor on the current CPU.
 969 *
 970 * Strangely, this function does important things even if it "fails",
 971 * which is especially common if the link is not up yet.  Hopefully
 972 * these things are all "harmless" if done twice!
 973 */
 974static void tile_net_register(void *dev_ptr)
 975{
 976	struct net_device *dev = (struct net_device *)dev_ptr;
 977	struct tile_net_priv *priv = netdev_priv(dev);
 978	int my_cpu = smp_processor_id();
 979	struct tile_net_cpu *info;
 980
 981	struct tile_netio_queue *queue;
 982
 983	/* Only network cpus can receive packets. */
 984	int queue_id =
 985		cpumask_test_cpu(my_cpu, &priv->network_cpus_map) ? 0 : 255;
 986
 987	netio_input_config_t config = {
 988		.flags = 0,
 989		.num_receive_packets = priv->network_cpus_credits,
 990		.queue_id = queue_id
 991	};
 992
 993	int ret = 0;
 994	netio_queue_impl_t *queuep;
 995
 996	PDEBUG("tile_net_register(queue_id %d)\n", queue_id);
 997
 998	if (!strcmp(dev->name, "xgbe0"))
 999		info = &__get_cpu_var(hv_xgbe0);
1000	else if (!strcmp(dev->name, "xgbe1"))
1001		info = &__get_cpu_var(hv_xgbe1);
1002	else if (!strcmp(dev->name, "gbe0"))
1003		info = &__get_cpu_var(hv_gbe0);
1004	else if (!strcmp(dev->name, "gbe1"))
1005		info = &__get_cpu_var(hv_gbe1);
1006	else
1007		BUG();
1008
1009	/* Initialize the egress timer. */
1010	init_timer(&info->egress_timer);
1011	info->egress_timer.data = (long)info;
1012	info->egress_timer.function = tile_net_handle_egress_timer;
1013
1014	u64_stats_init(&info->stats.syncp);
1015
1016	priv->cpu[my_cpu] = info;
1017
1018	/*
1019	 * Register ourselves with LIPP.  This does a lot of stuff,
1020	 * including invoking the LIPP registration code.
1021	 */
1022	ret = hv_dev_pwrite(priv->hv_devhdl, 0,
1023			    (HV_VirtAddr)&config,
1024			    sizeof(netio_input_config_t),
1025			    NETIO_IPP_INPUT_REGISTER_OFF);
1026	PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1027	       ret);
1028	if (ret < 0) {
1029		if (ret != NETIO_LINK_DOWN) {
1030			printk(KERN_DEBUG "hv_dev_pwrite "
1031			       "NETIO_IPP_INPUT_REGISTER_OFF failure %d\n",
1032			       ret);
1033		}
1034		info->link_down = (ret == NETIO_LINK_DOWN);
1035		return;
1036	}
1037
1038	/*
1039	 * Get the pointer to our queue's system part.
1040	 */
1041
1042	ret = hv_dev_pread(priv->hv_devhdl, 0,
1043			   (HV_VirtAddr)&queuep,
1044			   sizeof(netio_queue_impl_t *),
1045			   NETIO_IPP_INPUT_REGISTER_OFF);
1046	PDEBUG("hv_dev_pread(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1047	       ret);
1048	PDEBUG("queuep %p\n", queuep);
1049	if (ret <= 0) {
1050		/* ISSUE: Shouldn't this be a fatal error? */
1051		pr_err("hv_dev_pread NETIO_IPP_INPUT_REGISTER_OFF failure\n");
1052		return;
1053	}
1054
1055	queue = &info->queue;
1056
1057	queue->__system_part = queuep;
1058
1059	memset(&queue->__user_part, 0, sizeof(netio_queue_user_impl_t));
1060
1061	/* This is traditionally "config.num_receive_packets / 2". */
1062	queue->__user_part.__receive_credit_interval = 4;
1063	queue->__user_part.__receive_credit_remaining =
1064		queue->__user_part.__receive_credit_interval;
1065
1066	/*
1067	 * Get a fastio index from the hypervisor.
1068	 * ISSUE: Shouldn't this check the result?
1069	 */
1070	ret = hv_dev_pread(priv->hv_devhdl, 0,
1071			   (HV_VirtAddr)&queue->__user_part.__fastio_index,
1072			   sizeof(queue->__user_part.__fastio_index),
1073			   NETIO_IPP_GET_FASTIO_OFF);
1074	PDEBUG("hv_dev_pread(NETIO_IPP_GET_FASTIO_OFF) returned %d\n", ret);
1075
1076	/* Now we are registered. */
1077	info->registered = true;
1078}
1079
1080
1081/*
1082 * Deregister with hypervisor on the current CPU.
1083 *
1084 * This simply discards all our credits, so no more packets will be
1085 * delivered to this tile.  There may still be packets in our queue.
1086 *
1087 * Also, disable the ingress interrupt.
1088 */
1089static void tile_net_deregister(void *dev_ptr)
1090{
1091	struct net_device *dev = (struct net_device *)dev_ptr;
1092	struct tile_net_priv *priv = netdev_priv(dev);
1093	int my_cpu = smp_processor_id();
1094	struct tile_net_cpu *info = priv->cpu[my_cpu];
1095
1096	/* Disable the ingress interrupt. */
1097	disable_percpu_irq(priv->intr_id);
1098
1099	/* Do nothing else if not registered. */
1100	if (info == NULL || !info->registered)
1101		return;
1102
1103	{
1104		struct tile_netio_queue *queue = &info->queue;
1105		netio_queue_user_impl_t *qup = &queue->__user_part;
1106
1107		/* Discard all our credits. */
1108		__netio_fastio_return_credits(qup->__fastio_index, -1);
1109	}
1110}
1111
1112
1113/*
1114 * Unregister with hypervisor on the current CPU.
1115 *
1116 * Also, disable the ingress interrupt.
1117 */
1118static void tile_net_unregister(void *dev_ptr)
1119{
1120	struct net_device *dev = (struct net_device *)dev_ptr;
1121	struct tile_net_priv *priv = netdev_priv(dev);
1122	int my_cpu = smp_processor_id();
1123	struct tile_net_cpu *info = priv->cpu[my_cpu];
1124
1125	int ret;
1126	int dummy = 0;
1127
1128	/* Disable the ingress interrupt. */
1129	disable_percpu_irq(priv->intr_id);
1130
1131	/* Do nothing else if not registered. */
1132	if (info == NULL || !info->registered)
1133		return;
1134
1135	/* Unregister ourselves with LIPP/LEPP. */
1136	ret = hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1137			    sizeof(dummy), NETIO_IPP_INPUT_UNREGISTER_OFF);
1138	if (ret < 0)
1139		panic("Failed to unregister with LIPP/LEPP!\n");
1140
1141	/* Discard all packets still in our NetIO queue. */
1142	tile_net_discard_packets(dev);
1143
1144	/* Reset state. */
1145	info->num_needed_small_buffers = 0;
1146	info->num_needed_large_buffers = 0;
1147
1148	/* Cancel egress timer. */
1149	del_timer(&info->egress_timer);
1150	info->egress_timer_scheduled = false;
1151}
1152
1153
1154/*
1155 * Helper function for "tile_net_stop()".
1156 *
1157 * Also used to handle registration failure in "tile_net_open_inner()",
1158 * when the various extra steps in "tile_net_stop()" are not necessary.
1159 */
1160static void tile_net_stop_aux(struct net_device *dev)
1161{
1162	struct tile_net_priv *priv = netdev_priv(dev);
1163	int i;
1164
1165	int dummy = 0;
1166
1167	/*
1168	 * Unregister all tiles, so LIPP will stop delivering packets.
1169	 * Also, delete all the "napi" objects (sequentially, to protect
1170	 * "dev->napi_list").
1171	 */
1172	on_each_cpu(tile_net_unregister, (void *)dev, 1);
1173	for_each_online_cpu(i) {
1174		struct tile_net_cpu *info = priv->cpu[i];
1175		if (info != NULL && info->registered) {
1176			netif_napi_del(&info->napi);
1177			info->registered = false;
1178		}
1179	}
1180
1181	/* Stop LIPP/LEPP. */
1182	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1183			  sizeof(dummy), NETIO_IPP_STOP_SHIM_OFF) < 0)
1184		panic("Failed to stop LIPP/LEPP!\n");
1185
1186	priv->partly_opened = false;
1187}
1188
1189
1190/*
1191 * Disable NAPI for the given device on the current cpu.
1192 */
1193static void tile_net_stop_disable(void *dev_ptr)
1194{
1195	struct net_device *dev = (struct net_device *)dev_ptr;
1196	struct tile_net_priv *priv = netdev_priv(dev);
1197	int my_cpu = smp_processor_id();
1198	struct tile_net_cpu *info = priv->cpu[my_cpu];
1199
1200	/* Disable NAPI if needed. */
1201	if (info != NULL && info->napi_enabled) {
1202		napi_disable(&info->napi);
1203		info->napi_enabled = false;
1204	}
1205}
1206
1207
1208/*
1209 * Enable NAPI and the ingress interrupt for the given device
1210 * on the current cpu.
1211 *
1212 * ISSUE: Only do this for "network cpus"?
1213 */
1214static void tile_net_open_enable(void *dev_ptr)
1215{
1216	struct net_device *dev = (struct net_device *)dev_ptr;
1217	struct tile_net_priv *priv = netdev_priv(dev);
1218	int my_cpu = smp_processor_id();
1219	struct tile_net_cpu *info = priv->cpu[my_cpu];
1220
1221	/* Enable NAPI. */
1222	napi_enable(&info->napi);
1223	info->napi_enabled = true;
1224
1225	/* Enable the ingress interrupt. */
1226	enable_percpu_irq(priv->intr_id, 0);
1227}
1228
1229
1230/*
1231 * tile_net_open_inner does most of the work of bringing up the interface.
1232 * It's called from tile_net_open(), and also from tile_net_retry_open().
1233 * The return value is 0 if the interface was brought up, < 0 if
1234 * tile_net_open() should return the return value as an error, and > 0 if
1235 * tile_net_open() should return success and schedule a work item to
1236 * periodically retry the bringup.
1237 */
1238static int tile_net_open_inner(struct net_device *dev)
1239{
1240	struct tile_net_priv *priv = netdev_priv(dev);
1241	int my_cpu = smp_processor_id();
1242	struct tile_net_cpu *info;
1243	struct tile_netio_queue *queue;
1244	int result = 0;
1245	int i;
1246	int dummy = 0;
1247
1248	/*
1249	 * First try to register just on the local CPU, and handle any
1250	 * semi-expected "link down" failure specially.  Note that we
1251	 * do NOT call "tile_net_stop_aux()", unlike below.
1252	 */
1253	tile_net_register(dev);
1254	info = priv->cpu[my_cpu];
1255	if (!info->registered) {
1256		if (info->link_down)
1257			return 1;
1258		return -EAGAIN;
1259	}
1260
1261	/*
1262	 * Now register everywhere else.  If any registration fails,
1263	 * even for "link down" (which might not be possible), we
1264	 * clean up using "tile_net_stop_aux()".  Also, add all the
1265	 * "napi" objects (sequentially, to protect "dev->napi_list").
1266	 * ISSUE: Only use "netif_napi_add()" for "network cpus"?
1267	 */
1268	smp_call_function(tile_net_register, (void *)dev, 1);
1269	for_each_online_cpu(i) {
1270		struct tile_net_cpu *info = priv->cpu[i];
1271		if (info->registered)
1272			netif_napi_add(dev, &info->napi, tile_net_poll, 64);
1273		else
1274			result = -EAGAIN;
1275	}
1276	if (result != 0) {
1277		tile_net_stop_aux(dev);
1278		return result;
1279	}
1280
1281	queue = &info->queue;
1282
1283	if (priv->intr_id == 0) {
1284		unsigned int irq;
1285
1286		/*
1287		 * Acquire the irq allocated by the hypervisor.  Every
1288		 * queue gets the same irq.  The "__intr_id" field is
1289		 * "1 << irq", so we use "__ffs()" to extract "irq".
1290		 */
1291		priv->intr_id = queue->__system_part->__intr_id;
1292		BUG_ON(priv->intr_id == 0);
1293		irq = __ffs(priv->intr_id);
1294
1295		/*
1296		 * Register the ingress interrupt handler for this
1297		 * device, permanently.
1298		 *
1299		 * We used to call "free_irq()" in "tile_net_stop()",
1300		 * and then re-register the handler here every time,
1301		 * but that caused DNP errors in "handle_IRQ_event()"
1302		 * because "desc->action" was NULL.  See bug 9143.
1303		 */
1304		tile_irq_activate(irq, TILE_IRQ_PERCPU);
1305		BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
1306				   0, dev->name, (void *)dev) != 0);
1307	}
1308
1309	{
1310		/* Allocate initial buffers. */
1311
1312		int max_buffers =
1313			priv->network_cpus_count * priv->network_cpus_credits;
1314
1315		info->num_needed_small_buffers =
1316			min(LIPP_SMALL_BUFFERS, max_buffers);
1317
1318		info->num_needed_large_buffers =
1319			min(LIPP_LARGE_BUFFERS, max_buffers);
1320
1321		tile_net_provide_needed_buffers(info);
1322
1323		if (info->num_needed_small_buffers != 0 ||
1324		    info->num_needed_large_buffers != 0)
1325			panic("Insufficient memory for buffer stack!");
1326	}
1327
1328	/* We are about to be active. */
1329	priv->active = true;
1330
1331	/* Make sure "active" is visible to all tiles. */
1332	mb();
1333
1334	/* On each tile, enable NAPI and the ingress interrupt. */
1335	on_each_cpu(tile_net_open_enable, (void *)dev, 1);
1336
1337	/* Start LIPP/LEPP and activate "ingress" at the shim. */
1338	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1339			  sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
1340		panic("Failed to activate the LIPP Shim!\n");
1341
1342	/* Start our transmit queue. */
1343	netif_start_queue(dev);
1344
1345	return 0;
1346}
1347
1348
1349/*
1350 * Called periodically to retry bringing up the NetIO interface,
1351 * if it doesn't come up cleanly during tile_net_open().
1352 */
1353static void tile_net_open_retry(struct work_struct *w)
1354{
1355	struct delayed_work *dw =
1356		container_of(w, struct delayed_work, work);
1357
1358	struct tile_net_priv *priv =
1359		container_of(dw, struct tile_net_priv, retry_work);
1360
1361	/*
1362	 * Try to bring the NetIO interface up.  If it fails, reschedule
1363	 * ourselves to try again later; otherwise, tell Linux we now have
1364	 * a working link.  ISSUE: What if the return value is negative?
1365	 */
1366	if (tile_net_open_inner(priv->dev) != 0)
1367		schedule_delayed_work(&priv->retry_work,
1368				      TILE_NET_RETRY_INTERVAL);
1369	else
1370		netif_carrier_on(priv->dev);
1371}
1372
1373
1374/*
1375 * Called when a network interface is made active.
1376 *
1377 * Returns 0 on success, negative value on failure.
1378 *
1379 * The open entry point is called when a network interface is made
1380 * active by the system (IFF_UP).  At this point all resources needed
1381 * for transmit and receive operations are allocated, the interrupt
1382 * handler is registered with the OS (if needed), the watchdog timer
1383 * is started, and the stack is notified that the interface is ready.
1384 *
1385 * If the actual link is not available yet, then we tell Linux that
1386 * we have no carrier, and we keep checking until the link comes up.
1387 */
1388static int tile_net_open(struct net_device *dev)
1389{
1390	int ret = 0;
1391	struct tile_net_priv *priv = netdev_priv(dev);
1392
1393	/*
1394	 * We rely on priv->partly_opened to tell us if this is the
1395	 * first time this interface is being brought up. If it is
1396	 * set, the IPP was already initialized and should not be
1397	 * initialized again.
1398	 */
1399	if (!priv->partly_opened) {
1400
1401		int count;
1402		int credits;
1403
1404		/* Initialize LIPP/LEPP, and start the Shim. */
1405		ret = tile_net_open_aux(dev);
1406		if (ret < 0) {
1407			pr_err("tile_net_open_aux failed: %d\n", ret);
1408			return ret;
1409		}
1410
1411		/* Analyze the network cpus. */
1412
1413		if (network_cpus_used)
1414			cpumask_copy(&priv->network_cpus_map,
1415				     &network_cpus_map);
1416		else
1417			cpumask_copy(&priv->network_cpus_map, cpu_online_mask);
1418
1419
1420		count = cpumask_weight(&priv->network_cpus_map);
1421
1422		/* Limit credits to available buffers, and apply min. */
1423		credits = max(16, (LIPP_LARGE_BUFFERS / count) & ~1);
1424
1425		/* Apply "GBE" max limit. */
1426		/* ISSUE: Use higher limit for XGBE? */
1427		credits = min(NETIO_MAX_RECEIVE_PKTS, credits);
1428
1429		priv->network_cpus_count = count;
1430		priv->network_cpus_credits = credits;
1431
1432#ifdef TILE_NET_DEBUG
1433		pr_info("Using %d network cpus, with %d credits each\n",
1434		       priv->network_cpus_count, priv->network_cpus_credits);
1435#endif
1436
1437		priv->partly_opened = true;
1438
1439	} else {
1440		/* FIXME: Is this possible? */
1441		/* printk("Already partly opened.\n"); */
1442	}
1443
1444	/*
1445	 * Attempt to bring up the link.
1446	 */
1447	ret = tile_net_open_inner(dev);
1448	if (ret <= 0) {
1449		if (ret == 0)
1450			netif_carrier_on(dev);
1451		return ret;
1452	}
1453
1454	/*
1455	 * We were unable to bring up the NetIO interface, but we want to
1456	 * try again in a little bit.  Tell Linux that we have no carrier
1457	 * so it doesn't try to use the interface before the link comes up
1458	 * and then remember to try again later.
1459	 */
1460	netif_carrier_off(dev);
1461	schedule_delayed_work(&priv->retry_work, TILE_NET_RETRY_INTERVAL);
1462
1463	return 0;
1464}
1465
1466
1467static int tile_net_drain_lipp_buffers(struct tile_net_priv *priv)
1468{
1469	int n = 0;
1470
1471	/* Drain all the LIPP buffers. */
1472	while (true) {
1473		unsigned int buffer;
1474
1475		/* NOTE: This should never fail. */
1476		if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&buffer,
1477				 sizeof(buffer), NETIO_IPP_DRAIN_OFF) < 0)
1478			break;
1479
1480		/* Stop when done. */
1481		if (buffer == 0)
1482			break;
1483
1484		{
1485			/* Convert "linux_buffer_t" to "va". */
1486			void *va = __va((phys_addr_t)(buffer >> 1) << 7);
1487
1488			/* Acquire the associated "skb". */
1489			struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
1490			struct sk_buff *skb = *skb_ptr;
1491
1492			kfree_skb(skb);
1493		}
1494
1495		n++;
1496	}
1497
1498	return n;
1499}
1500
1501
1502/*
1503 * Disables a network interface.
1504 *
1505 * Returns 0, this is not allowed to fail.
1506 *
1507 * The close entry point is called when an interface is de-activated
1508 * by the OS.  The hardware is still under the drivers control, but
1509 * needs to be disabled.  A global MAC reset is issued to stop the
1510 * hardware, and all transmit and receive resources are freed.
1511 *
1512 * ISSUE: How closely does "netif_running(dev)" mirror "priv->active"?
1513 *
1514 * Before we are called by "__dev_close()", "netif_running()" will
1515 * have been cleared, so no NEW calls to "tile_net_poll()" will be
1516 * made by "netpoll_poll_dev()".
1517 *
1518 * Often, this can cause some tiles to still have packets in their
1519 * queues, so we must call "tile_net_discard_packets()" later.
1520 *
1521 * Note that some other tile may still be INSIDE "tile_net_poll()",
1522 * and in fact, many will be, if there is heavy network load.
1523 *
1524 * Calling "on_each_cpu(tile_net_stop_disable, (void *)dev, 1)" when
1525 * any tile is still "napi_schedule()"'d will induce a horrible crash
1526 * when "msleep()" is called.  This includes tiles which are inside
1527 * "tile_net_poll()" which have not yet called "napi_complete()".
1528 *
1529 * So, we must first try to wait long enough for other tiles to finish
1530 * with any current "tile_net_poll()" call, and, hopefully, to clear
1531 * the "scheduled" flag.  ISSUE: It is unclear what happens to tiles
1532 * which have called "napi_schedule()" but which had not yet tried to
1533 * call "tile_net_poll()", or which exhausted their budget inside
1534 * "tile_net_poll()" just before this function was called.
1535 */
1536static int tile_net_stop(struct net_device *dev)
1537{
1538	struct tile_net_priv *priv = netdev_priv(dev);
1539
1540	PDEBUG("tile_net_stop()\n");
1541
1542	/* Start discarding packets. */
1543	priv->active = false;
1544
1545	/* Make sure "active" is visible to all tiles. */
1546	mb();
1547
1548	/*
1549	 * On each tile, make sure no NEW packets get delivered, and
1550	 * disable the ingress interrupt.
1551	 *
1552	 * Note that the ingress interrupt can fire AFTER this,
1553	 * presumably due to packets which were recently delivered,
1554	 * but it will have no effect.
1555	 */
1556	on_each_cpu(tile_net_deregister, (void *)dev, 1);
1557
1558	/* Optimistically drain LIPP buffers. */
1559	(void)tile_net_drain_lipp_buffers(priv);
1560
1561	/* ISSUE: Only needed if not yet fully open. */
1562	cancel_delayed_work_sync(&priv->retry_work);
1563
1564	/* Can't transmit any more. */
1565	netif_stop_queue(dev);
1566
1567	/* Disable NAPI on each tile. */
1568	on_each_cpu(tile_net_stop_disable, (void *)dev, 1);
1569
1570	/*
1571	 * Drain any remaining LIPP buffers.  NOTE: This "printk()"
1572	 * has never been observed, but in theory it could happen.
1573	 */
1574	if (tile_net_drain_lipp_buffers(priv) != 0)
1575		printk("Had to drain some extra LIPP buffers!\n");
1576
1577	/* Stop LIPP/LEPP. */
1578	tile_net_stop_aux(dev);
1579
1580	/*
1581	 * ISSUE: It appears that, in practice anyway, by the time we
1582	 * get here, there are no pending completions, but just in case,
1583	 * we free (all of) them anyway.
1584	 */
1585	while (tile_net_lepp_free_comps(dev, true))
1586		/* loop */;
1587
1588	/* Wipe the EPP queue, and wait till the stores hit the EPP. */
1589	memset(priv->eq, 0, sizeof(lepp_queue_t));
1590	mb();
1591
1592	return 0;
1593}
1594
1595
1596/*
1597 * Prepare the "frags" info for the resulting LEPP command.
1598 *
1599 * If needed, flush the memory used by the frags.
1600 */
1601static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
1602				      struct sk_buff *skb,
1603				      void *b_data, unsigned int b_len)
1604{
1605	unsigned int i, n = 0;
1606
1607	struct skb_shared_info *sh = skb_shinfo(skb);
1608
1609	phys_addr_t cpa;
1610
1611	if (b_len != 0) {
1612
1613		if (!hash_default)
1614			finv_buffer_remote(b_data, b_len, 0);
1615
1616		cpa = __pa(b_data);
1617		frags[n].cpa_lo = cpa;
1618		frags[n].cpa_hi = cpa >> 32;
1619		frags[n].length = b_len;
1620		frags[n].hash_for_home = hash_default;
1621		n++;
1622	}
1623
1624	for (i = 0; i < sh->nr_frags; i++) {
1625
1626		skb_frag_t *f = &sh->frags[i];
1627		unsigned long pfn = page_to_pfn(skb_frag_page(f));
1628
1629		/* FIXME: Compute "hash_for_home" properly. */
1630		/* ISSUE: The hypervisor checks CHIP_HAS_REV1_DMA_PACKETS(). */
1631		int hash_for_home = hash_default;
1632
1633		/* FIXME: Hmmm. */
1634		if (!hash_default) {
1635			void *va = pfn_to_kaddr(pfn) + f->page_offset;
1636			BUG_ON(PageHighMem(skb_frag_page(f)));
1637			finv_buffer_remote(va, skb_frag_size(f), 0);
1638		}
1639
1640		cpa = ((phys_addr_t)pfn << PAGE_SHIFT) + f->page_offset;
1641		frags[n].cpa_lo = cpa;
1642		frags[n].cpa_hi = cpa >> 32;
1643		frags[n].length = skb_frag_size(f);
1644		frags[n].hash_for_home = hash_for_home;
1645		n++;
1646	}
1647
1648	return n;
1649}
1650
1651
1652/*
1653 * This function takes "skb", consisting of a header template and a
1654 * payload, and hands it to LEPP, to emit as one or more segments,
1655 * each consisting of a possibly modified header, plus a piece of the
1656 * payload, via a process known as "tcp segmentation offload".
1657 *
1658 * Usually, "data" will contain the header template, of size "sh_len",
1659 * and "sh->frags" will contain "skb->data_len" bytes of payload, and
1660 * there will be "sh->gso_segs" segments.
1661 *
1662 * Sometimes, if "sendfile()" requires copying, we will be called with
1663 * "data" containing the header and payload, with "frags" being empty.
1664 *
1665 * Sometimes, for example when using NFS over TCP, a single segment can
1666 * span 3 fragments, which must be handled carefully in LEPP.
1667 *
1668 * See "emulate_large_send_offload()" for some reference code, which
1669 * does not handle checksumming.
1670 *
1671 * ISSUE: How do we make sure that high memory DMA does not migrate?
1672 */
1673static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
1674{
1675	struct tile_net_priv *priv = netdev_priv(dev);
1676	int my_cpu = smp_processor_id();
1677	struct tile_net_cpu *info = priv->cpu[my_cpu];
1678	struct tile_net_stats_t *stats = &info->stats;
1679
1680	struct skb_shared_info *sh = skb_shinfo(skb);
1681
1682	unsigned char *data = skb->data;
1683
1684	/* The ip header follows the ethernet header. */
1685	struct iphdr *ih = ip_hdr(skb);
1686	unsigned int ih_len = ih->ihl * 4;
1687
1688	/* Note that "nh == ih", by definition. */
1689	unsigned char *nh = skb_network_header(skb);
1690	unsigned int eh_len = nh - data;
1691
1692	/* The tcp header follows the ip header. */
1693	struct tcphdr *th = (struct tcphdr *)(nh + ih_len);
1694	unsigned int th_len = th->doff * 4;
1695
1696	/* The total number of header bytes. */
1697	/* NOTE: This may be less than skb_headlen(skb). */
1698	unsigned int sh_len = eh_len + ih_len + th_len;
1699
1700	/* The number of payload bytes at "skb->data + sh_len". */
1701	/* This is non-zero for sendfile() without HIGHDMA. */
1702	unsigned int b_len = skb_headlen(skb) - sh_len;
1703
1704	/* The total number of payload bytes. */
1705	unsigned int d_len = b_len + skb->data_len;
1706
1707	/* The maximum payload size. */
1708	unsigned int p_len = sh->gso_size;
1709
1710	/* The total number of segments. */
1711	unsigned int num_segs = sh->gso_segs;
1712
1713	/* The temporary copy of the command. */
1714	u32 cmd_body[(LEPP_MAX_CMD_SIZE + 3) / 4];
1715	lepp_tso_cmd_t *cmd = (lepp_tso_cmd_t *)cmd_body;
1716
1717	/* Analyze the "frags". */
1718	unsigned int num_frags =
1719		tile_net_tx_frags(cmd->frags, skb, data + sh_len, b_len);
1720
1721	/* The size of the command, including frags and header. */
1722	size_t cmd_size = LEPP_TSO_CMD_SIZE(num_frags, sh_len);
1723
1724	/* The command header. */
1725	lepp_tso_cmd_t cmd_init = {
1726		.tso = true,
1727		.header_size = sh_len,
1728		.ip_offset = eh_len,
1729		.tcp_offset = eh_len + ih_len,
1730		.payload_size = p_len,
1731		.num_frags = num_frags,
1732	};
1733
1734	unsigned long irqflags;
1735
1736	lepp_queue_t *eq = priv->eq;
1737
1738	struct sk_buff *olds[8];
1739	unsigned int wanted = 8;
1740	unsigned int i, nolds = 0;
1741
1742	unsigned int cmd_head, cmd_tail, cmd_next;
1743	unsigned int comp_tail;
1744
1745
1746	/* Paranoia. */
1747	BUG_ON(skb->protocol != htons(ETH_P_IP));
1748	BUG_ON(ih->protocol != IPPROTO_TCP);
1749	BUG_ON(skb->ip_summed != CHECKSUM_PARTIAL);
1750	BUG_ON(num_frags > LEPP_MAX_FRAGS);
1751	/*--BUG_ON(num_segs != (d_len + (p_len - 1)) / p_len); */
1752	BUG_ON(num_segs <= 1);
1753
1754
1755	/* Finish preparing the command. */
1756
1757	/* Copy the command header. */
1758	*cmd = cmd_init;
1759
1760	/* Copy the "header". */
1761	memcpy(&cmd->frags[num_frags], data, sh_len);
1762
1763
1764	/* Prefetch and wait, to minimize time spent holding the spinlock. */
1765	prefetch_L1(&eq->comp_tail);
1766	prefetch_L1(&eq->cmd_tail);
1767	mb();
1768
1769
1770	/* Enqueue the command. */
1771
1772	spin_lock_irqsave(&priv->eq_lock, irqflags);
1773
1774	/* Handle completions if needed to make room. */
1775	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
1776	if (lepp_num_free_comp_slots(eq) == 0) {
1777		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
1778		if (nolds == 0) {
1779busy:
1780			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1781			return NETDEV_TX_BUSY;
1782		}
1783	}
1784
1785	cmd_head = eq->cmd_head;
1786	cmd_tail = eq->cmd_tail;
1787
1788	/* Prepare to advance, detecting full queue. */
1789	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
1790	cmd_next = cmd_tail + cmd_size;
1791	if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1792		goto busy;
1793	if (cmd_next > LEPP_CMD_LIMIT) {
1794		cmd_next = 0;
1795		if (cmd_next == cmd_head)
1796			goto busy;
1797	}
1798
1799	/* Copy the command. */
1800	memcpy(&eq->cmds[cmd_tail], cmd, cmd_size);
1801
1802	/* Advance. */
1803	cmd_tail = cmd_next;
1804
1805	/* Record "skb" for eventual freeing. */
1806	comp_tail = eq->comp_tail;
1807	eq->comps[comp_tail] = skb;
1808	LEPP_QINC(comp_tail);
1809	eq->comp_tail = comp_tail;
1810
1811	/* Flush before allowing LEPP to handle the command. */
1812	/* ISSUE: Is this the optimal location for the flush? */
1813	__insn_mf();
1814
1815	eq->cmd_tail = cmd_tail;
1816
1817	/* NOTE: Using "4" here is more efficient than "0" or "2", */
1818	/* and, strangely, more efficient than pre-checking the number */
1819	/* of available completions, and comparing it to 4. */
1820	if (nolds == 0)
1821		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
1822
1823	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1824
1825	/* Handle completions. */
1826	for (i = 0; i < nolds; i++)
1827		dev_consume_skb_any(olds[i]);
1828
1829	/* Update stats. */
1830	u64_stats_update_begin(&stats->syncp);
1831	stats->tx_packets += num_segs;
1832	stats->tx_bytes += (num_segs * sh_len) + d_len;
1833	u64_stats_update_end(&stats->syncp);
1834
1835	/* Make sure the egress timer is scheduled. */
1836	tile_net_schedule_egress_timer(info);
1837
1838	return NETDEV_TX_OK;
1839}
1840
1841
1842/*
1843 * Transmit a packet (called by the kernel via "hard_start_xmit" hook).
1844 */
1845static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
1846{
1847	struct tile_net_priv *priv = netdev_priv(dev);
1848	int my_cpu = smp_processor_id();
1849	struct tile_net_cpu *info = priv->cpu[my_cpu];
1850	struct tile_net_stats_t *stats = &info->stats;
1851
1852	unsigned long irqflags;
1853
1854	struct skb_shared_info *sh = skb_shinfo(skb);
1855
1856	unsigned int len = skb->len;
1857	unsigned char *data = skb->data;
1858
1859	unsigned int csum_start = skb_checksum_start_offset(skb);
1860
1861	lepp_frag_t frags[1 + MAX_SKB_FRAGS];
1862
1863	unsigned int num_frags;
1864
1865	lepp_queue_t *eq = priv->eq;
1866
1867	struct sk_buff *olds[8];
1868	unsigned int wanted = 8;
1869	unsigned int i, nolds = 0;
1870
1871	unsigned int cmd_size = sizeof(lepp_cmd_t);
1872
1873	unsigned int cmd_head, cmd_tail, cmd_next;
1874	unsigned int comp_tail;
1875
1876	lepp_cmd_t cmds[1 + MAX_SKB_FRAGS];
1877
1878
1879	/*
1880	 * This is paranoia, since we think that if the link doesn't come
1881	 * up, telling Linux we have no carrier will keep it from trying
1882	 * to transmit.  If it does, though, we can't execute this routine,
1883	 * since data structures we depend on aren't set up yet.
1884	 */
1885	if (!info->registered)
1886		return NETDEV_TX_BUSY;
1887
1888
1889	/* Save the timestamp. */
1890	dev->trans_start = jiffies;
1891
1892
1893#ifdef TILE_NET_PARANOIA
1894#if CHIP_HAS_CBOX_HOME_MAP()
1895	if (hash_default) {
1896		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)data);
1897		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
1898			panic("Non-HFH egress buffer! VA=%p Mode=%d PTE=%llx",
1899			      data, hv_pte_get_mode(pte), hv_pte_val(pte));
1900	}
1901#endif
1902#endif
1903
1904
1905#ifdef TILE_NET_DUMP_PACKETS
1906	/* ISSUE: Does not dump the "frags". */
1907	dump_packet(data, skb_headlen(skb), "tx");
1908#endif /* TILE_NET_DUMP_PACKETS */
1909
1910
1911	if (sh->gso_size != 0)
1912		return tile_net_tx_tso(skb, dev);
1913
1914
1915	/* Prepare the commands. */
1916
1917	num_frags = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
1918
1919	for (i = 0; i < num_frags; i++) {
1920
1921		bool final = (i == num_frags - 1);
1922
1923		lepp_cmd_t cmd = {
1924			.cpa_lo = frags[i].cpa_lo,
1925			.cpa_hi = frags[i].cpa_hi,
1926			.length = frags[i].length,
1927			.hash_for_home = frags[i].hash_for_home,
1928			.send_completion = final,
1929			.end_of_packet = final
1930		};
1931
1932		if (i == 0 && skb->ip_summed == CHECKSUM_PARTIAL) {
1933			cmd.compute_checksum = 1;
1934			cmd.checksum_data.bits.start_byte = csum_start;
1935			cmd.checksum_data.bits.count = len - csum_start;
1936			cmd.checksum_data.bits.destination_byte =
1937				csum_start + skb->csum_offset;
1938		}
1939
1940		cmds[i] = cmd;
1941	}
1942
1943
1944	/* Prefetch and wait, to minimize time spent holding the spinlock. */
1945	prefetch_L1(&eq->comp_tail);
1946	prefetch_L1(&eq->cmd_tail);
1947	mb();
1948
1949
1950	/* Enqueue the commands. */
1951
1952	spin_lock_irqsave(&priv->eq_lock, irqflags);
1953
1954	/* Handle completions if needed to make room. */
1955	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
1956	if (lepp_num_free_comp_slots(eq) == 0) {
1957		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
1958		if (nolds == 0) {
1959busy:
1960			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1961			return NETDEV_TX_BUSY;
1962		}
1963	}
1964
1965	cmd_head = eq->cmd_head;
1966	cmd_tail = eq->cmd_tail;
1967
1968	/* Copy the commands, or fail. */
1969	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
1970	for (i = 0; i < num_frags; i++) {
1971
1972		/* Prepare to advance, detecting full queue. */
1973		cmd_next = cmd_tail + cmd_size;
1974		if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1975			goto busy;
1976		if (cmd_next > LEPP_CMD_LIMIT) {
1977			cmd_next = 0;
1978			if (cmd_next == cmd_head)
1979				goto busy;
1980		}
1981
1982		/* Copy the command. */
1983		*(lepp_cmd_t *)&eq->cmds[cmd_tail] = cmds[i];
1984
1985		/* Advance. */
1986		cmd_tail = cmd_next;
1987	}
1988
1989	/* Record "skb" for eventual freeing. */
1990	comp_tail = eq->comp_tail;
1991	eq->comps[comp_tail] = skb;
1992	LEPP_QINC(comp_tail);
1993	eq->comp_tail = comp_tail;
1994
1995	/* Flush before allowing LEPP to handle the command. */
1996	/* ISSUE: Is this the optimal location for the flush? */
1997	__insn_mf();
1998
1999	eq->cmd_tail = cmd_tail;
2000
2001	/* NOTE: Using "4" here is more efficient than "0" or "2", */
2002	/* and, strangely, more efficient than pre-checking the number */
2003	/* of available completions, and comparing it to 4. */
2004	if (nolds == 0)
2005		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
2006
2007	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
2008
2009	/* Handle completions. */
2010	for (i = 0; i < nolds; i++)
2011		dev_consume_skb_any(olds[i]);
2012
2013	/* HACK: Track "expanded" size for short packets (e.g. 42 < 60). */
2014	u64_stats_update_begin(&stats->syncp);
2015	stats->tx_packets++;
2016	stats->tx_bytes += ((len >= ETH_ZLEN) ? len : ETH_ZLEN);
2017	u64_stats_update_end(&stats->syncp);
2018
2019	/* Make sure the egress timer is scheduled. */
2020	tile_net_schedule_egress_timer(info);
2021
2022	return NETDEV_TX_OK;
2023}
2024
2025
2026/*
2027 * Deal with a transmit timeout.
2028 */
2029static void tile_net_tx_timeout(struct net_device *dev)
2030{
2031	PDEBUG("tile_net_tx_timeout()\n");
2032	PDEBUG("Transmit timeout at %ld, latency %ld\n", jiffies,
2033	       jiffies - dev->trans_start);
2034
2035	/* XXX: ISSUE: This doesn't seem useful for us. */
2036	netif_wake_queue(dev);
2037}
2038
2039
2040/*
2041 * Ioctl commands.
2042 */
2043static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2044{
2045	return -EOPNOTSUPP;
2046}
2047
2048
2049/*
2050 * Get System Network Statistics.
2051 *
2052 * Returns the address of the device statistics structure.
2053 */
2054static struct rtnl_link_stats64 *tile_net_get_stats64(struct net_device *dev,
2055		struct rtnl_link_stats64 *stats)
2056{
2057	struct tile_net_priv *priv = netdev_priv(dev);
2058	u64 rx_packets = 0, tx_packets = 0;
2059	u64 rx_bytes = 0, tx_bytes = 0;
2060	u64 rx_errors = 0, rx_dropped = 0;
2061	int i;
2062
2063	for_each_online_cpu(i) {
2064		struct tile_net_stats_t *cpu_stats;
2065		u64 trx_packets, ttx_packets, trx_bytes, ttx_bytes;
2066		u64 trx_errors, trx_dropped;
2067		unsigned int start;
2068
2069		if (priv->cpu[i] == NULL)
2070			continue;
2071		cpu_stats = &priv->cpu[i]->stats;
2072
2073		do {
2074			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
2075			trx_packets = cpu_stats->rx_packets;
2076			ttx_packets = cpu_stats->tx_packets;
2077			trx_bytes   = cpu_stats->rx_bytes;
2078			ttx_bytes   = cpu_stats->tx_bytes;
2079			trx_errors  = cpu_stats->rx_errors;
2080			trx_dropped = cpu_stats->rx_dropped;
2081		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
2082
2083		rx_packets += trx_packets;
2084		tx_packets += ttx_packets;
2085		rx_bytes   += trx_bytes;
2086		tx_bytes   += ttx_bytes;
2087		rx_errors  += trx_errors;
2088		rx_dropped += trx_dropped;
2089	}
2090
2091	stats->rx_packets = rx_packets;
2092	stats->tx_packets = tx_packets;
2093	stats->rx_bytes   = rx_bytes;
2094	stats->tx_bytes   = tx_bytes;
2095	stats->rx_errors  = rx_errors;
2096	stats->rx_dropped = rx_dropped;
2097
2098	return stats;
2099}
2100
2101
2102/*
2103 * Change the "mtu".
2104 *
2105 * The "change_mtu" method is usually not needed.
2106 * If you need it, it must be like this.
2107 */
2108static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
2109{
2110	PDEBUG("tile_net_change_mtu()\n");
2111
2112	/* Check ranges. */
2113	if ((new_mtu < 68) || (new_mtu > 1500))
2114		return -EINVAL;
2115
2116	/* Accept the value. */
2117	dev->mtu = new_mtu;
2118
2119	return 0;
2120}
2121
2122
2123/*
2124 * Change the Ethernet Address of the NIC.
2125 *
2126 * The hypervisor driver does not support changing MAC address.  However,
2127 * the IPP does not do anything with the MAC address, so the address which
2128 * gets used on outgoing packets, and which is accepted on incoming packets,
2129 * is completely up to the NetIO program or kernel driver which is actually
2130 * handling them.
2131 *
2132 * Returns 0 on success, negative on failure.
2133 */
2134static int tile_net_set_mac_address(struct net_device *dev, void *p)
2135{
2136	struct sockaddr *addr = p;
2137
2138	if (!is_valid_ether_addr(addr->sa_data))
2139		return -EADDRNOTAVAIL;
2140
2141	/* ISSUE: Note that "dev_addr" is now a pointer. */
2142	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
2143
2144	return 0;
2145}
2146
2147
2148/*
2149 * Obtain the MAC address from the hypervisor.
2150 * This must be done before opening the device.
2151 */
2152static int tile_net_get_mac(struct net_device *dev)
2153{
2154	struct tile_net_priv *priv = netdev_priv(dev);
2155
2156	char hv_dev_name[32];
2157	int len;
2158
2159	__netio_getset_offset_t offset = { .word = NETIO_IPP_PARAM_OFF };
2160
2161	int ret;
2162
2163	/* For example, "xgbe0". */
2164	strcpy(hv_dev_name, dev->name);
2165	len = strlen(hv_dev_name);
2166
2167	/* For example, "xgbe/0". */
2168	hv_dev_name[len] = hv_dev_name[len - 1];
2169	hv_dev_name[len - 1] = '/';
2170	len++;
2171
2172	/* For example, "xgbe/0/native_hash". */
2173	strcpy(hv_dev_name + len, hash_default ? "/native_hash" : "/native");
2174
2175	/* Get the hypervisor handle for this device. */
2176	priv->hv_devhdl = hv_dev_open((HV_VirtAddr)hv_dev_name, 0);
2177	PDEBUG("hv_dev_open(%s) returned %d %p\n",
2178	       hv_dev_name, priv->hv_devhdl, &priv->hv_devhdl);
2179	if (priv->hv_devhdl < 0) {
2180		if (priv->hv_devhdl == HV_ENODEV)
2181			printk(KERN_DEBUG "Ignoring unconfigured device %s\n",
2182				 hv_dev_name);
2183		else
2184			printk(KERN_DEBUG "hv_dev_open(%s) returned %d\n",
2185				 hv_dev_name, priv->hv_devhdl);
2186		return -1;
2187	}
2188
2189	/*
2190	 * Read the hardware address from the hypervisor.
2191	 * ISSUE: Note that "dev_addr" is now a pointer.
2192	 */
2193	offset.bits.class = NETIO_PARAM;
2194	offset.bits.addr = NETIO_PARAM_MAC;
2195	ret = hv_dev_pread(priv->hv_devhdl, 0,
2196			   (HV_VirtAddr)dev->dev_addr, dev->addr_len,
2197			   offset.word);
2198	PDEBUG("hv_dev_pread(NETIO_PARAM_MAC) returned %d\n", ret);
2199	if (ret <= 0) {
2200		printk(KERN_DEBUG "hv_dev_pread(NETIO_PARAM_MAC) %s failed\n",
2201		       dev->name);
2202		/*
2203		 * Since the device is configured by the hypervisor but we
2204		 * can't get its MAC address, we are most likely running
2205		 * the simulator, so let's generate a random MAC address.
2206		 */
2207		eth_hw_addr_random(dev);
2208	}
2209
2210	return 0;
2211}
2212
2213
2214#ifdef CONFIG_NET_POLL_CONTROLLER
2215/*
2216 * Polling 'interrupt' - used by things like netconsole to send skbs
2217 * without having to re-enable interrupts. It's not called while
2218 * the interrupt routine is executing.
2219 */
2220static void tile_net_netpoll(struct net_device *dev)
2221{
2222	struct tile_net_priv *priv = netdev_priv(dev);
2223	disable_percpu_irq(priv->intr_id);
2224	tile_net_handle_ingress_interrupt(priv->intr_id, dev);
2225	enable_percpu_irq(priv->intr_id, 0);
2226}
2227#endif
2228
2229
2230static const struct net_device_ops tile_net_ops = {
2231	.ndo_open = tile_net_open,
2232	.ndo_stop = tile_net_stop,
2233	.ndo_start_xmit = tile_net_tx,
2234	.ndo_do_ioctl = tile_net_ioctl,
2235	.ndo_get_stats64 = tile_net_get_stats64,
2236	.ndo_change_mtu = tile_net_change_mtu,
2237	.ndo_tx_timeout = tile_net_tx_timeout,
2238	.ndo_set_mac_address = tile_net_set_mac_address,
2239#ifdef CONFIG_NET_POLL_CONTROLLER
2240	.ndo_poll_controller = tile_net_netpoll,
2241#endif
2242};
2243
2244
2245/*
2246 * The setup function.
2247 *
2248 * This uses ether_setup() to assign various fields in dev, including
2249 * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
2250 */
2251static void tile_net_setup(struct net_device *dev)
2252{
2253	netdev_features_t features = 0;
2254
2255	ether_setup(dev);
2256	dev->netdev_ops = &tile_net_ops;
2257	dev->watchdog_timeo = TILE_NET_TIMEOUT;
2258	dev->tx_queue_len = TILE_NET_TX_QUEUE_LEN;
2259	dev->mtu = TILE_NET_MTU;
2260
2261	features |= NETIF_F_HW_CSUM;
2262	features |= NETIF_F_SG;
2263
2264	/* We support TSO iff the HV supports sufficient frags. */
2265	if (LEPP_MAX_FRAGS >= 1 + MAX_SKB_FRAGS)
2266		features |= NETIF_F_TSO;
2267
2268	/* We can't support HIGHDMA without hash_default, since we need
2269	 * to be able to finv() with a VA if we don't have hash_default.
2270	 */
2271	if (hash_default)
2272		features |= NETIF_F_HIGHDMA;
2273
2274	dev->hw_features   |= features;
2275	dev->vlan_features |= features;
2276	dev->features      |= features;
2277}
2278
2279
2280/*
2281 * Allocate the device structure, register the device, and obtain the
2282 * MAC address from the hypervisor.
2283 */
2284static struct net_device *tile_net_dev_init(const char *name)
2285{
2286	int ret;
2287	struct net_device *dev;
2288	struct tile_net_priv *priv;
2289
2290	/*
2291	 * Allocate the device structure.  This allocates "priv", calls
2292	 * tile_net_setup(), and saves "name".  Normally, "name" is a
2293	 * template, instantiated by register_netdev(), but not for us.
2294	 */
2295	dev = alloc_netdev(sizeof(*priv), name, tile_net_setup);
2296	if (!dev) {
2297		pr_err("alloc_netdev(%s) failed\n", name);
2298		return NULL;
2299	}
2300
2301	priv = netdev_priv(dev);
2302
2303	/* Initialize "priv". */
2304
2305	memset(priv, 0, sizeof(*priv));
2306
2307	/* Save "dev" for "tile_net_open_retry()". */
2308	priv->dev = dev;
2309
2310	INIT_DELAYED_WORK(&priv->retry_work, tile_net_open_retry);
2311
2312	spin_lock_init(&priv->eq_lock);
2313
2314	/* Allocate "eq". */
2315	priv->eq_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, EQ_ORDER);
2316	if (!priv->eq_pages) {
2317		free_netdev(dev);
2318		return NULL;
2319	}
2320	priv->eq = page_address(priv->eq_pages);
2321
2322	/* Register the network device. */
2323	ret = register_netdev(dev);
2324	if (ret) {
2325		pr_err("register_netdev %s failed %d\n", dev->name, ret);
2326		__free_pages(priv->eq_pages, EQ_ORDER);
2327		free_netdev(dev);
2328		return NULL;
2329	}
2330
2331	/* Get the MAC address. */
2332	ret = tile_net_get_mac(dev);
2333	if (ret < 0) {
2334		unregister_netdev(dev);
2335		__free_pages(priv->eq_pages, EQ_ORDER);
2336		free_netdev(dev);
2337		return NULL;
2338	}
2339
2340	return dev;
2341}
2342
2343
2344/*
2345 * Module cleanup.
2346 *
2347 * FIXME: If compiled as a module, this module cannot be "unloaded",
2348 * because the "ingress interrupt handler" is registered permanently.
2349 */
2350static void tile_net_cleanup(void)
2351{
2352	int i;
2353
2354	for (i = 0; i < TILE_NET_DEVS; i++) {
2355		if (tile_net_devs[i]) {
2356			struct net_device *dev = tile_net_devs[i];
2357			struct tile_net_priv *priv = netdev_priv(dev);
2358			unregister_netdev(dev);
2359			finv_buffer_remote(priv->eq, EQ_SIZE, 0);
2360			__free_pages(priv->eq_pages, EQ_ORDER);
2361			free_netdev(dev);
2362		}
2363	}
2364}
2365
2366
2367/*
2368 * Module initialization.
2369 */
2370static int tile_net_init_module(void)
2371{
2372	pr_info("Tilera Network Driver\n");
2373
2374	tile_net_devs[0] = tile_net_dev_init("xgbe0");
2375	tile_net_devs[1] = tile_net_dev_init("xgbe1");
2376	tile_net_devs[2] = tile_net_dev_init("gbe0");
2377	tile_net_devs[3] = tile_net_dev_init("gbe1");
2378
2379	return 0;
2380}
2381
2382
2383module_init(tile_net_init_module);
2384module_exit(tile_net_cleanup);
2385
2386
2387#ifndef MODULE
2388
2389/*
2390 * The "network_cpus" boot argument specifies the cpus that are dedicated
2391 * to handle ingress packets.
2392 *
2393 * The parameter should be in the form "network_cpus=m-n[,x-y]", where
2394 * m, n, x, y are integer numbers that represent the cpus that can be
2395 * neither a dedicated cpu nor a dataplane cpu.
2396 */
2397static int __init network_cpus_setup(char *str)
2398{
2399	int rc = cpulist_parse_crop(str, &network_cpus_map);
2400	if (rc != 0) {
2401		pr_warning("network_cpus=%s: malformed cpu list\n",
2402		       str);
2403	} else {
2404
2405		/* Remove dedicated cpus. */
2406		cpumask_and(&network_cpus_map, &network_cpus_map,
2407			    cpu_possible_mask);
2408
2409
2410		if (cpumask_empty(&network_cpus_map)) {
2411			pr_warning("Ignoring network_cpus='%s'.\n",
2412			       str);
2413		} else {
2414			char buf[1024];
2415			cpulist_scnprintf(buf, sizeof(buf), &network_cpus_map);
2416			pr_info("Linux network CPUs: %s\n", buf);
2417			network_cpus_used = true;
2418		}
2419	}
2420
2421	return 0;
2422}
2423__setup("network_cpus=", network_cpus_setup);
2424
2425#endif