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 ETH_DATA_LEN
 
 
 
 
 
 
 
  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(&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#ifdef TILE_NET_DUMP_PACKETS
 725	dump_packet(buf, len, "rx");
 726#endif /* TILE_NET_DUMP_PACKETS */
 727
 728#ifdef TILE_NET_VERIFY_INGRESS
 729	if (pkt_status == NETIO_PKT_STATUS_OVERSIZE && len >= 64) {
 
 
 
 
 
 
 730		dump_packet(buf, len, "rx");
 731		panic("Unexpected OVERSIZE.");
 
 
 
 
 
 
 
 
 
 
 732	}
 733#endif
 734
 735	filter = 0;
 736
 737	if (pkt_status == NETIO_PKT_STATUS_BAD) {
 738		/* Handle CRC error and hardware truncation. */
 739		filter = 2;
 740	} else if (!(dev->flags & IFF_UP)) {
 741		/* Filter packets received before we're up. */
 742		filter = 1;
 743	} else if (NETIO_PKT_ETHERTYPE_RECOGNIZED_M(metadata, pkt) &&
 744		   pkt_status == NETIO_PKT_STATUS_UNDERSIZE) {
 745		/* Filter "truncated" packets. */
 746		filter = 2;
 747	} else if (!(dev->flags & IFF_PROMISC)) {
 
 748		if (!is_multicast_ether_addr(buf)) {
 749			/* Filter packets not for our address. */
 750			const u8 *mine = dev->dev_addr;
 751			filter = !ether_addr_equal(mine, buf);
 752		}
 753	}
 754
 755	u64_stats_update_begin(&stats->syncp);
 756
 757	if (filter != 0) {
 758
 759		if (filter == 1)
 760			stats->rx_dropped++;
 761		else
 762			stats->rx_errors++;
 763
 764		tile_net_provide_linux_buffer(info, va, small);
 765
 766	} else {
 767
 768		/* Acquire the associated "skb". */
 769		struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
 770		struct sk_buff *skb = *skb_ptr;
 771
 772		/* Paranoia. */
 773		if (skb->data != buf)
 774			panic("Corrupt linux buffer from LIPP! "
 775			      "VA=%p, skb=%p, skb->data=%p\n",
 776			      va, skb, skb->data);
 777
 778		/* Encode the actual packet length. */
 779		skb_put(skb, len);
 780
 781		/* NOTE: This call also sets "skb->dev = dev". */
 782		skb->protocol = eth_type_trans(skb, dev);
 783
 784		/* Avoid recomputing "good" TCP/UDP checksums. */
 785		if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
 786			skb->ip_summed = CHECKSUM_UNNECESSARY;
 787
 788		netif_receive_skb(skb);
 789
 790		stats->rx_packets++;
 791		stats->rx_bytes += len;
 792	}
 793
 794	u64_stats_update_end(&stats->syncp);
 795
 796	/* ISSUE: It would be nice to defer this until the packet has */
 797	/* actually been processed. */
 798	tile_net_return_credit(info);
 799
 800	/* Consume this packet. */
 801	qup->__packet_receive_read = index2;
 802
 803	return !filter;
 804}
 805
 806
 807/*
 808 * Handle some packets for the given device on the current CPU.
 809 *
 810 * If "tile_net_stop()" is called on some other tile while this
 811 * function is running, we will return, hopefully before that
 812 * other tile asks us to call "napi_disable()".
 813 *
 814 * The "rotting packet" race condition occurs if a packet arrives
 815 * during the extremely narrow window between the queue appearing to
 816 * be empty, and the ingress interrupt being re-enabled.  This happens
 817 * a LOT under heavy network load.
 818 */
 819static int tile_net_poll(struct napi_struct *napi, int budget)
 820{
 821	struct net_device *dev = napi->dev;
 822	struct tile_net_priv *priv = netdev_priv(dev);
 823	int my_cpu = smp_processor_id();
 824	struct tile_net_cpu *info = priv->cpu[my_cpu];
 825	struct tile_netio_queue *queue = &info->queue;
 826	netio_queue_impl_t *qsp = queue->__system_part;
 827	netio_queue_user_impl_t *qup = &queue->__user_part;
 828
 829	unsigned int work = 0;
 830
 831	if (budget <= 0)
 832		goto done;
 833
 834	while (priv->active) {
 835		int index = qup->__packet_receive_read;
 836		if (index == qsp->__packet_receive_queue.__packet_write)
 837			break;
 838
 839		if (tile_net_poll_aux(info, index)) {
 840			if (++work >= budget)
 841				goto done;
 842		}
 843	}
 844
 845	napi_complete(&info->napi);
 846
 847	if (!priv->active)
 848		goto done;
 849
 850	/* Re-enable the ingress interrupt. */
 851	enable_percpu_irq(priv->intr_id, 0);
 852
 853	/* HACK: Avoid the "rotting packet" problem (see above). */
 854	if (qup->__packet_receive_read !=
 855	    qsp->__packet_receive_queue.__packet_write) {
 856		/* ISSUE: Sometimes this returns zero, presumably */
 857		/* because an interrupt was handled for this tile. */
 858		(void)napi_reschedule(&info->napi);
 859	}
 860
 861done:
 862
 863	if (priv->active)
 864		tile_net_provide_needed_buffers(info);
 865
 866	return work;
 867}
 868
 869
 870/*
 871 * Handle an ingress interrupt for the given device on the current cpu.
 872 *
 873 * ISSUE: Sometimes this gets called after "disable_percpu_irq()" has
 874 * been called!  This is probably due to "pending hypervisor downcalls".
 875 *
 876 * ISSUE: Is there any race condition between the "napi_schedule()" here
 877 * and the "napi_complete()" call above?
 878 */
 879static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
 880{
 881	struct net_device *dev = (struct net_device *)dev_ptr;
 882	struct tile_net_priv *priv = netdev_priv(dev);
 883	int my_cpu = smp_processor_id();
 884	struct tile_net_cpu *info = priv->cpu[my_cpu];
 885
 886	/* Disable the ingress interrupt. */
 887	disable_percpu_irq(priv->intr_id);
 888
 889	/* Ignore unwanted interrupts. */
 890	if (!priv->active)
 891		return IRQ_HANDLED;
 892
 893	/* ISSUE: Sometimes "info->napi_enabled" is false here. */
 894
 895	napi_schedule(&info->napi);
 896
 897	return IRQ_HANDLED;
 898}
 899
 900
 901/*
 902 * One time initialization per interface.
 903 */
 904static int tile_net_open_aux(struct net_device *dev)
 905{
 906	struct tile_net_priv *priv = netdev_priv(dev);
 907
 908	int ret;
 909	int dummy;
 910	unsigned int epp_lotar;
 911
 912	/*
 913	 * Find out where EPP memory should be homed.
 914	 */
 915	ret = hv_dev_pread(priv->hv_devhdl, 0,
 916			   (HV_VirtAddr)&epp_lotar, sizeof(epp_lotar),
 917			   NETIO_EPP_SHM_OFF);
 918	if (ret < 0) {
 919		pr_err("could not read epp_shm_queue lotar.\n");
 920		return -EIO;
 921	}
 922
 923	/*
 924	 * Home the page on the EPP.
 925	 */
 926	{
 927		int epp_home = hv_lotar_to_cpu(epp_lotar);
 928		homecache_change_page_home(priv->eq_pages, EQ_ORDER, epp_home);
 929	}
 930
 931	/*
 932	 * Register the EPP shared memory queue.
 933	 */
 934	{
 935		netio_ipp_address_t ea = {
 936			.va = 0,
 937			.pa = __pa(priv->eq),
 938			.pte = hv_pte(0),
 939			.size = EQ_SIZE,
 940		};
 941		ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
 942		ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
 943		ret = hv_dev_pwrite(priv->hv_devhdl, 0,
 944				    (HV_VirtAddr)&ea,
 945				    sizeof(ea),
 946				    NETIO_EPP_SHM_OFF);
 947		if (ret < 0)
 948			return -EIO;
 949	}
 950
 951	/*
 952	 * Start LIPP/LEPP.
 953	 */
 954	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
 955			  sizeof(dummy), NETIO_IPP_START_SHIM_OFF) < 0) {
 956		pr_warn("Failed to start LIPP/LEPP\n");
 957		return -EIO;
 958	}
 959
 960	return 0;
 961}
 962
 963
 964/*
 965 * Register with hypervisor on the current CPU.
 966 *
 967 * Strangely, this function does important things even if it "fails",
 968 * which is especially common if the link is not up yet.  Hopefully
 969 * these things are all "harmless" if done twice!
 970 */
 971static void tile_net_register(void *dev_ptr)
 972{
 973	struct net_device *dev = (struct net_device *)dev_ptr;
 974	struct tile_net_priv *priv = netdev_priv(dev);
 975	int my_cpu = smp_processor_id();
 976	struct tile_net_cpu *info;
 977
 978	struct tile_netio_queue *queue;
 979
 980	/* Only network cpus can receive packets. */
 981	int queue_id =
 982		cpumask_test_cpu(my_cpu, &priv->network_cpus_map) ? 0 : 255;
 983
 984	netio_input_config_t config = {
 985		.flags = 0,
 986		.num_receive_packets = priv->network_cpus_credits,
 987		.queue_id = queue_id
 988	};
 989
 990	int ret = 0;
 991	netio_queue_impl_t *queuep;
 992
 993	PDEBUG("tile_net_register(queue_id %d)\n", queue_id);
 994
 995	if (!strcmp(dev->name, "xgbe0"))
 996		info = this_cpu_ptr(&hv_xgbe0);
 997	else if (!strcmp(dev->name, "xgbe1"))
 998		info = this_cpu_ptr(&hv_xgbe1);
 999	else if (!strcmp(dev->name, "gbe0"))
1000		info = this_cpu_ptr(&hv_gbe0);
1001	else if (!strcmp(dev->name, "gbe1"))
1002		info = this_cpu_ptr(&hv_gbe1);
1003	else
1004		BUG();
1005
1006	/* Initialize the egress timer. */
1007	init_timer_pinned(&info->egress_timer);
1008	info->egress_timer.data = (long)info;
1009	info->egress_timer.function = tile_net_handle_egress_timer;
1010
1011	u64_stats_init(&info->stats.syncp);
1012
1013	priv->cpu[my_cpu] = info;
1014
1015	/*
1016	 * Register ourselves with LIPP.  This does a lot of stuff,
1017	 * including invoking the LIPP registration code.
1018	 */
1019	ret = hv_dev_pwrite(priv->hv_devhdl, 0,
1020			    (HV_VirtAddr)&config,
1021			    sizeof(netio_input_config_t),
1022			    NETIO_IPP_INPUT_REGISTER_OFF);
1023	PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1024	       ret);
1025	if (ret < 0) {
1026		if (ret != NETIO_LINK_DOWN) {
1027			printk(KERN_DEBUG "hv_dev_pwrite "
1028			       "NETIO_IPP_INPUT_REGISTER_OFF failure %d\n",
1029			       ret);
1030		}
1031		info->link_down = (ret == NETIO_LINK_DOWN);
1032		return;
1033	}
1034
1035	/*
1036	 * Get the pointer to our queue's system part.
1037	 */
1038
1039	ret = hv_dev_pread(priv->hv_devhdl, 0,
1040			   (HV_VirtAddr)&queuep,
1041			   sizeof(netio_queue_impl_t *),
1042			   NETIO_IPP_INPUT_REGISTER_OFF);
1043	PDEBUG("hv_dev_pread(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
1044	       ret);
1045	PDEBUG("queuep %p\n", queuep);
1046	if (ret <= 0) {
1047		/* ISSUE: Shouldn't this be a fatal error? */
1048		pr_err("hv_dev_pread NETIO_IPP_INPUT_REGISTER_OFF failure\n");
1049		return;
1050	}
1051
1052	queue = &info->queue;
1053
1054	queue->__system_part = queuep;
1055
1056	memset(&queue->__user_part, 0, sizeof(netio_queue_user_impl_t));
1057
1058	/* This is traditionally "config.num_receive_packets / 2". */
1059	queue->__user_part.__receive_credit_interval = 4;
1060	queue->__user_part.__receive_credit_remaining =
1061		queue->__user_part.__receive_credit_interval;
1062
1063	/*
1064	 * Get a fastio index from the hypervisor.
1065	 * ISSUE: Shouldn't this check the result?
1066	 */
1067	ret = hv_dev_pread(priv->hv_devhdl, 0,
1068			   (HV_VirtAddr)&queue->__user_part.__fastio_index,
1069			   sizeof(queue->__user_part.__fastio_index),
1070			   NETIO_IPP_GET_FASTIO_OFF);
1071	PDEBUG("hv_dev_pread(NETIO_IPP_GET_FASTIO_OFF) returned %d\n", ret);
1072
1073	/* Now we are registered. */
1074	info->registered = true;
1075}
1076
1077
1078/*
1079 * Deregister with hypervisor on the current CPU.
1080 *
1081 * This simply discards all our credits, so no more packets will be
1082 * delivered to this tile.  There may still be packets in our queue.
1083 *
1084 * Also, disable the ingress interrupt.
1085 */
1086static void tile_net_deregister(void *dev_ptr)
1087{
1088	struct net_device *dev = (struct net_device *)dev_ptr;
1089	struct tile_net_priv *priv = netdev_priv(dev);
1090	int my_cpu = smp_processor_id();
1091	struct tile_net_cpu *info = priv->cpu[my_cpu];
1092
1093	/* Disable the ingress interrupt. */
1094	disable_percpu_irq(priv->intr_id);
1095
1096	/* Do nothing else if not registered. */
1097	if (info == NULL || !info->registered)
1098		return;
1099
1100	{
1101		struct tile_netio_queue *queue = &info->queue;
1102		netio_queue_user_impl_t *qup = &queue->__user_part;
1103
1104		/* Discard all our credits. */
1105		__netio_fastio_return_credits(qup->__fastio_index, -1);
1106	}
1107}
1108
1109
1110/*
1111 * Unregister with hypervisor on the current CPU.
1112 *
1113 * Also, disable the ingress interrupt.
1114 */
1115static void tile_net_unregister(void *dev_ptr)
1116{
1117	struct net_device *dev = (struct net_device *)dev_ptr;
1118	struct tile_net_priv *priv = netdev_priv(dev);
1119	int my_cpu = smp_processor_id();
1120	struct tile_net_cpu *info = priv->cpu[my_cpu];
1121
1122	int ret;
1123	int dummy = 0;
1124
1125	/* Disable the ingress interrupt. */
1126	disable_percpu_irq(priv->intr_id);
1127
1128	/* Do nothing else if not registered. */
1129	if (info == NULL || !info->registered)
1130		return;
1131
1132	/* Unregister ourselves with LIPP/LEPP. */
1133	ret = hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1134			    sizeof(dummy), NETIO_IPP_INPUT_UNREGISTER_OFF);
1135	if (ret < 0)
1136		panic("Failed to unregister with LIPP/LEPP!\n");
1137
1138	/* Discard all packets still in our NetIO queue. */
1139	tile_net_discard_packets(dev);
1140
1141	/* Reset state. */
1142	info->num_needed_small_buffers = 0;
1143	info->num_needed_large_buffers = 0;
1144
1145	/* Cancel egress timer. */
1146	del_timer(&info->egress_timer);
1147	info->egress_timer_scheduled = false;
1148}
1149
1150
1151/*
1152 * Helper function for "tile_net_stop()".
1153 *
1154 * Also used to handle registration failure in "tile_net_open_inner()",
1155 * when the various extra steps in "tile_net_stop()" are not necessary.
1156 */
1157static void tile_net_stop_aux(struct net_device *dev)
1158{
1159	struct tile_net_priv *priv = netdev_priv(dev);
1160	int i;
1161
1162	int dummy = 0;
1163
1164	/*
1165	 * Unregister all tiles, so LIPP will stop delivering packets.
1166	 * Also, delete all the "napi" objects (sequentially, to protect
1167	 * "dev->napi_list").
1168	 */
1169	on_each_cpu(tile_net_unregister, (void *)dev, 1);
1170	for_each_online_cpu(i) {
1171		struct tile_net_cpu *info = priv->cpu[i];
1172		if (info != NULL && info->registered) {
1173			netif_napi_del(&info->napi);
1174			info->registered = false;
1175		}
1176	}
1177
1178	/* Stop LIPP/LEPP. */
1179	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1180			  sizeof(dummy), NETIO_IPP_STOP_SHIM_OFF) < 0)
1181		panic("Failed to stop LIPP/LEPP!\n");
1182
1183	priv->partly_opened = false;
1184}
1185
1186
1187/*
1188 * Disable NAPI for the given device on the current cpu.
1189 */
1190static void tile_net_stop_disable(void *dev_ptr)
1191{
1192	struct net_device *dev = (struct net_device *)dev_ptr;
1193	struct tile_net_priv *priv = netdev_priv(dev);
1194	int my_cpu = smp_processor_id();
1195	struct tile_net_cpu *info = priv->cpu[my_cpu];
1196
1197	/* Disable NAPI if needed. */
1198	if (info != NULL && info->napi_enabled) {
1199		napi_disable(&info->napi);
1200		info->napi_enabled = false;
1201	}
1202}
1203
1204
1205/*
1206 * Enable NAPI and the ingress interrupt for the given device
1207 * on the current cpu.
1208 *
1209 * ISSUE: Only do this for "network cpus"?
1210 */
1211static void tile_net_open_enable(void *dev_ptr)
1212{
1213	struct net_device *dev = (struct net_device *)dev_ptr;
1214	struct tile_net_priv *priv = netdev_priv(dev);
1215	int my_cpu = smp_processor_id();
1216	struct tile_net_cpu *info = priv->cpu[my_cpu];
1217
1218	/* Enable NAPI. */
1219	napi_enable(&info->napi);
1220	info->napi_enabled = true;
1221
1222	/* Enable the ingress interrupt. */
1223	enable_percpu_irq(priv->intr_id, 0);
1224}
1225
1226
1227/*
1228 * tile_net_open_inner does most of the work of bringing up the interface.
1229 * It's called from tile_net_open(), and also from tile_net_retry_open().
1230 * The return value is 0 if the interface was brought up, < 0 if
1231 * tile_net_open() should return the return value as an error, and > 0 if
1232 * tile_net_open() should return success and schedule a work item to
1233 * periodically retry the bringup.
1234 */
1235static int tile_net_open_inner(struct net_device *dev)
1236{
1237	struct tile_net_priv *priv = netdev_priv(dev);
1238	int my_cpu = smp_processor_id();
1239	struct tile_net_cpu *info;
1240	struct tile_netio_queue *queue;
1241	int result = 0;
1242	int i;
1243	int dummy = 0;
1244
1245	/*
1246	 * First try to register just on the local CPU, and handle any
1247	 * semi-expected "link down" failure specially.  Note that we
1248	 * do NOT call "tile_net_stop_aux()", unlike below.
1249	 */
1250	tile_net_register(dev);
1251	info = priv->cpu[my_cpu];
1252	if (!info->registered) {
1253		if (info->link_down)
1254			return 1;
1255		return -EAGAIN;
1256	}
1257
1258	/*
1259	 * Now register everywhere else.  If any registration fails,
1260	 * even for "link down" (which might not be possible), we
1261	 * clean up using "tile_net_stop_aux()".  Also, add all the
1262	 * "napi" objects (sequentially, to protect "dev->napi_list").
1263	 * ISSUE: Only use "netif_napi_add()" for "network cpus"?
1264	 */
1265	smp_call_function(tile_net_register, (void *)dev, 1);
1266	for_each_online_cpu(i) {
1267		struct tile_net_cpu *info = priv->cpu[i];
1268		if (info->registered)
1269			netif_napi_add(dev, &info->napi, tile_net_poll, 64);
1270		else
1271			result = -EAGAIN;
1272	}
1273	if (result != 0) {
1274		tile_net_stop_aux(dev);
1275		return result;
1276	}
1277
1278	queue = &info->queue;
1279
1280	if (priv->intr_id == 0) {
1281		unsigned int irq;
1282
1283		/*
1284		 * Acquire the irq allocated by the hypervisor.  Every
1285		 * queue gets the same irq.  The "__intr_id" field is
1286		 * "1 << irq", so we use "__ffs()" to extract "irq".
1287		 */
1288		priv->intr_id = queue->__system_part->__intr_id;
1289		BUG_ON(priv->intr_id == 0);
1290		irq = __ffs(priv->intr_id);
1291
1292		/*
1293		 * Register the ingress interrupt handler for this
1294		 * device, permanently.
1295		 *
1296		 * We used to call "free_irq()" in "tile_net_stop()",
1297		 * and then re-register the handler here every time,
1298		 * but that caused DNP errors in "handle_IRQ_event()"
1299		 * because "desc->action" was NULL.  See bug 9143.
1300		 */
1301		tile_irq_activate(irq, TILE_IRQ_PERCPU);
1302		BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
1303				   0, dev->name, (void *)dev) != 0);
1304	}
1305
1306	{
1307		/* Allocate initial buffers. */
1308
1309		int max_buffers =
1310			priv->network_cpus_count * priv->network_cpus_credits;
1311
1312		info->num_needed_small_buffers =
1313			min(LIPP_SMALL_BUFFERS, max_buffers);
1314
1315		info->num_needed_large_buffers =
1316			min(LIPP_LARGE_BUFFERS, max_buffers);
1317
1318		tile_net_provide_needed_buffers(info);
1319
1320		if (info->num_needed_small_buffers != 0 ||
1321		    info->num_needed_large_buffers != 0)
1322			panic("Insufficient memory for buffer stack!");
1323	}
1324
1325	/* We are about to be active. */
1326	priv->active = true;
1327
1328	/* Make sure "active" is visible to all tiles. */
1329	mb();
1330
1331	/* On each tile, enable NAPI and the ingress interrupt. */
1332	on_each_cpu(tile_net_open_enable, (void *)dev, 1);
1333
1334	/* Start LIPP/LEPP and activate "ingress" at the shim. */
1335	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
1336			  sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
1337		panic("Failed to activate the LIPP Shim!\n");
1338
1339	/* Start our transmit queue. */
1340	netif_start_queue(dev);
1341
1342	return 0;
1343}
1344
1345
1346/*
1347 * Called periodically to retry bringing up the NetIO interface,
1348 * if it doesn't come up cleanly during tile_net_open().
1349 */
1350static void tile_net_open_retry(struct work_struct *w)
1351{
1352	struct delayed_work *dw = to_delayed_work(w);
 
1353
1354	struct tile_net_priv *priv =
1355		container_of(dw, struct tile_net_priv, retry_work);
1356
1357	/*
1358	 * Try to bring the NetIO interface up.  If it fails, reschedule
1359	 * ourselves to try again later; otherwise, tell Linux we now have
1360	 * a working link.  ISSUE: What if the return value is negative?
1361	 */
1362	if (tile_net_open_inner(priv->dev) != 0)
1363		schedule_delayed_work(&priv->retry_work,
1364				      TILE_NET_RETRY_INTERVAL);
1365	else
1366		netif_carrier_on(priv->dev);
1367}
1368
1369
1370/*
1371 * Called when a network interface is made active.
1372 *
1373 * Returns 0 on success, negative value on failure.
1374 *
1375 * The open entry point is called when a network interface is made
1376 * active by the system (IFF_UP).  At this point all resources needed
1377 * for transmit and receive operations are allocated, the interrupt
1378 * handler is registered with the OS (if needed), the watchdog timer
1379 * is started, and the stack is notified that the interface is ready.
1380 *
1381 * If the actual link is not available yet, then we tell Linux that
1382 * we have no carrier, and we keep checking until the link comes up.
1383 */
1384static int tile_net_open(struct net_device *dev)
1385{
1386	int ret = 0;
1387	struct tile_net_priv *priv = netdev_priv(dev);
1388
1389	/*
1390	 * We rely on priv->partly_opened to tell us if this is the
1391	 * first time this interface is being brought up. If it is
1392	 * set, the IPP was already initialized and should not be
1393	 * initialized again.
1394	 */
1395	if (!priv->partly_opened) {
1396
1397		int count;
1398		int credits;
1399
1400		/* Initialize LIPP/LEPP, and start the Shim. */
1401		ret = tile_net_open_aux(dev);
1402		if (ret < 0) {
1403			pr_err("tile_net_open_aux failed: %d\n", ret);
1404			return ret;
1405		}
1406
1407		/* Analyze the network cpus. */
1408
1409		if (network_cpus_used)
1410			cpumask_copy(&priv->network_cpus_map,
1411				     &network_cpus_map);
1412		else
1413			cpumask_copy(&priv->network_cpus_map, cpu_online_mask);
1414
1415
1416		count = cpumask_weight(&priv->network_cpus_map);
1417
1418		/* Limit credits to available buffers, and apply min. */
1419		credits = max(16, (LIPP_LARGE_BUFFERS / count) & ~1);
1420
1421		/* Apply "GBE" max limit. */
1422		/* ISSUE: Use higher limit for XGBE? */
1423		credits = min(NETIO_MAX_RECEIVE_PKTS, credits);
1424
1425		priv->network_cpus_count = count;
1426		priv->network_cpus_credits = credits;
1427
1428#ifdef TILE_NET_DEBUG
1429		pr_info("Using %d network cpus, with %d credits each\n",
1430		       priv->network_cpus_count, priv->network_cpus_credits);
1431#endif
1432
1433		priv->partly_opened = true;
1434
1435	} else {
1436		/* FIXME: Is this possible? */
1437		/* printk("Already partly opened.\n"); */
1438	}
1439
1440	/*
1441	 * Attempt to bring up the link.
1442	 */
1443	ret = tile_net_open_inner(dev);
1444	if (ret <= 0) {
1445		if (ret == 0)
1446			netif_carrier_on(dev);
1447		return ret;
1448	}
1449
1450	/*
1451	 * We were unable to bring up the NetIO interface, but we want to
1452	 * try again in a little bit.  Tell Linux that we have no carrier
1453	 * so it doesn't try to use the interface before the link comes up
1454	 * and then remember to try again later.
1455	 */
1456	netif_carrier_off(dev);
1457	schedule_delayed_work(&priv->retry_work, TILE_NET_RETRY_INTERVAL);
1458
1459	return 0;
1460}
1461
1462
1463static int tile_net_drain_lipp_buffers(struct tile_net_priv *priv)
1464{
1465	int n = 0;
1466
1467	/* Drain all the LIPP buffers. */
1468	while (true) {
1469		unsigned int buffer;
1470
1471		/* NOTE: This should never fail. */
1472		if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&buffer,
1473				 sizeof(buffer), NETIO_IPP_DRAIN_OFF) < 0)
1474			break;
1475
1476		/* Stop when done. */
1477		if (buffer == 0)
1478			break;
1479
1480		{
1481			/* Convert "linux_buffer_t" to "va". */
1482			void *va = __va((phys_addr_t)(buffer >> 1) << 7);
1483
1484			/* Acquire the associated "skb". */
1485			struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
1486			struct sk_buff *skb = *skb_ptr;
1487
1488			kfree_skb(skb);
1489		}
1490
1491		n++;
1492	}
1493
1494	return n;
1495}
1496
1497
1498/*
1499 * Disables a network interface.
1500 *
1501 * Returns 0, this is not allowed to fail.
1502 *
1503 * The close entry point is called when an interface is de-activated
1504 * by the OS.  The hardware is still under the drivers control, but
1505 * needs to be disabled.  A global MAC reset is issued to stop the
1506 * hardware, and all transmit and receive resources are freed.
1507 *
1508 * ISSUE: How closely does "netif_running(dev)" mirror "priv->active"?
1509 *
1510 * Before we are called by "__dev_close()", "netif_running()" will
1511 * have been cleared, so no NEW calls to "tile_net_poll()" will be
1512 * made by "netpoll_poll_dev()".
1513 *
1514 * Often, this can cause some tiles to still have packets in their
1515 * queues, so we must call "tile_net_discard_packets()" later.
1516 *
1517 * Note that some other tile may still be INSIDE "tile_net_poll()",
1518 * and in fact, many will be, if there is heavy network load.
1519 *
1520 * Calling "on_each_cpu(tile_net_stop_disable, (void *)dev, 1)" when
1521 * any tile is still "napi_schedule()"'d will induce a horrible crash
1522 * when "msleep()" is called.  This includes tiles which are inside
1523 * "tile_net_poll()" which have not yet called "napi_complete()".
1524 *
1525 * So, we must first try to wait long enough for other tiles to finish
1526 * with any current "tile_net_poll()" call, and, hopefully, to clear
1527 * the "scheduled" flag.  ISSUE: It is unclear what happens to tiles
1528 * which have called "napi_schedule()" but which had not yet tried to
1529 * call "tile_net_poll()", or which exhausted their budget inside
1530 * "tile_net_poll()" just before this function was called.
1531 */
1532static int tile_net_stop(struct net_device *dev)
1533{
1534	struct tile_net_priv *priv = netdev_priv(dev);
1535
1536	PDEBUG("tile_net_stop()\n");
1537
1538	/* Start discarding packets. */
1539	priv->active = false;
1540
1541	/* Make sure "active" is visible to all tiles. */
1542	mb();
1543
1544	/*
1545	 * On each tile, make sure no NEW packets get delivered, and
1546	 * disable the ingress interrupt.
1547	 *
1548	 * Note that the ingress interrupt can fire AFTER this,
1549	 * presumably due to packets which were recently delivered,
1550	 * but it will have no effect.
1551	 */
1552	on_each_cpu(tile_net_deregister, (void *)dev, 1);
1553
1554	/* Optimistically drain LIPP buffers. */
1555	(void)tile_net_drain_lipp_buffers(priv);
1556
1557	/* ISSUE: Only needed if not yet fully open. */
1558	cancel_delayed_work_sync(&priv->retry_work);
1559
1560	/* Can't transmit any more. */
1561	netif_stop_queue(dev);
1562
1563	/* Disable NAPI on each tile. */
1564	on_each_cpu(tile_net_stop_disable, (void *)dev, 1);
1565
1566	/*
1567	 * Drain any remaining LIPP buffers.  NOTE: This "printk()"
1568	 * has never been observed, but in theory it could happen.
1569	 */
1570	if (tile_net_drain_lipp_buffers(priv) != 0)
1571		printk("Had to drain some extra LIPP buffers!\n");
1572
1573	/* Stop LIPP/LEPP. */
1574	tile_net_stop_aux(dev);
1575
1576	/*
1577	 * ISSUE: It appears that, in practice anyway, by the time we
1578	 * get here, there are no pending completions, but just in case,
1579	 * we free (all of) them anyway.
1580	 */
1581	while (tile_net_lepp_free_comps(dev, true))
1582		/* loop */;
1583
1584	/* Wipe the EPP queue, and wait till the stores hit the EPP. */
1585	memset(priv->eq, 0, sizeof(lepp_queue_t));
1586	mb();
1587
1588	return 0;
1589}
1590
1591
1592/*
1593 * Prepare the "frags" info for the resulting LEPP command.
1594 *
1595 * If needed, flush the memory used by the frags.
1596 */
1597static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
1598				      struct sk_buff *skb,
1599				      void *b_data, unsigned int b_len)
1600{
1601	unsigned int i, n = 0;
1602
1603	struct skb_shared_info *sh = skb_shinfo(skb);
1604
1605	phys_addr_t cpa;
1606
1607	if (b_len != 0) {
1608
1609		if (!hash_default)
1610			finv_buffer_remote(b_data, b_len, 0);
1611
1612		cpa = __pa(b_data);
1613		frags[n].cpa_lo = cpa;
1614		frags[n].cpa_hi = cpa >> 32;
1615		frags[n].length = b_len;
1616		frags[n].hash_for_home = hash_default;
1617		n++;
1618	}
1619
1620	for (i = 0; i < sh->nr_frags; i++) {
1621
1622		skb_frag_t *f = &sh->frags[i];
1623		unsigned long pfn = page_to_pfn(skb_frag_page(f));
1624
1625		/* FIXME: Compute "hash_for_home" properly. */
1626		/* ISSUE: The hypervisor checks CHIP_HAS_REV1_DMA_PACKETS(). */
1627		int hash_for_home = hash_default;
1628
1629		/* FIXME: Hmmm. */
1630		if (!hash_default) {
1631			void *va = pfn_to_kaddr(pfn) + f->page_offset;
1632			BUG_ON(PageHighMem(skb_frag_page(f)));
1633			finv_buffer_remote(va, skb_frag_size(f), 0);
1634		}
1635
1636		cpa = ((phys_addr_t)pfn << PAGE_SHIFT) + f->page_offset;
1637		frags[n].cpa_lo = cpa;
1638		frags[n].cpa_hi = cpa >> 32;
1639		frags[n].length = skb_frag_size(f);
1640		frags[n].hash_for_home = hash_for_home;
1641		n++;
1642	}
1643
1644	return n;
1645}
1646
1647
1648/*
1649 * This function takes "skb", consisting of a header template and a
1650 * payload, and hands it to LEPP, to emit as one or more segments,
1651 * each consisting of a possibly modified header, plus a piece of the
1652 * payload, via a process known as "tcp segmentation offload".
1653 *
1654 * Usually, "data" will contain the header template, of size "sh_len",
1655 * and "sh->frags" will contain "skb->data_len" bytes of payload, and
1656 * there will be "sh->gso_segs" segments.
1657 *
1658 * Sometimes, if "sendfile()" requires copying, we will be called with
1659 * "data" containing the header and payload, with "frags" being empty.
1660 *
1661 * Sometimes, for example when using NFS over TCP, a single segment can
1662 * span 3 fragments, which must be handled carefully in LEPP.
1663 *
1664 * See "emulate_large_send_offload()" for some reference code, which
1665 * does not handle checksumming.
1666 *
1667 * ISSUE: How do we make sure that high memory DMA does not migrate?
1668 */
1669static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
1670{
1671	struct tile_net_priv *priv = netdev_priv(dev);
1672	int my_cpu = smp_processor_id();
1673	struct tile_net_cpu *info = priv->cpu[my_cpu];
1674	struct tile_net_stats_t *stats = &info->stats;
1675
1676	struct skb_shared_info *sh = skb_shinfo(skb);
1677
1678	unsigned char *data = skb->data;
1679
1680	/* The ip header follows the ethernet header. */
1681	struct iphdr *ih = ip_hdr(skb);
1682	unsigned int ih_len = ih->ihl * 4;
1683
1684	/* Note that "nh == ih", by definition. */
1685	unsigned char *nh = skb_network_header(skb);
1686	unsigned int eh_len = nh - data;
1687
1688	/* The tcp header follows the ip header. */
1689	struct tcphdr *th = (struct tcphdr *)(nh + ih_len);
1690	unsigned int th_len = th->doff * 4;
1691
1692	/* The total number of header bytes. */
1693	/* NOTE: This may be less than skb_headlen(skb). */
1694	unsigned int sh_len = eh_len + ih_len + th_len;
1695
1696	/* The number of payload bytes at "skb->data + sh_len". */
1697	/* This is non-zero for sendfile() without HIGHDMA. */
1698	unsigned int b_len = skb_headlen(skb) - sh_len;
1699
1700	/* The total number of payload bytes. */
1701	unsigned int d_len = b_len + skb->data_len;
1702
1703	/* The maximum payload size. */
1704	unsigned int p_len = sh->gso_size;
1705
1706	/* The total number of segments. */
1707	unsigned int num_segs = sh->gso_segs;
1708
1709	/* The temporary copy of the command. */
1710	u32 cmd_body[(LEPP_MAX_CMD_SIZE + 3) / 4];
1711	lepp_tso_cmd_t *cmd = (lepp_tso_cmd_t *)cmd_body;
1712
1713	/* Analyze the "frags". */
1714	unsigned int num_frags =
1715		tile_net_tx_frags(cmd->frags, skb, data + sh_len, b_len);
1716
1717	/* The size of the command, including frags and header. */
1718	size_t cmd_size = LEPP_TSO_CMD_SIZE(num_frags, sh_len);
1719
1720	/* The command header. */
1721	lepp_tso_cmd_t cmd_init = {
1722		.tso = true,
1723		.header_size = sh_len,
1724		.ip_offset = eh_len,
1725		.tcp_offset = eh_len + ih_len,
1726		.payload_size = p_len,
1727		.num_frags = num_frags,
1728	};
1729
1730	unsigned long irqflags;
1731
1732	lepp_queue_t *eq = priv->eq;
1733
1734	struct sk_buff *olds[8];
1735	unsigned int wanted = 8;
1736	unsigned int i, nolds = 0;
1737
1738	unsigned int cmd_head, cmd_tail, cmd_next;
1739	unsigned int comp_tail;
1740
1741
1742	/* Paranoia. */
1743	BUG_ON(skb->protocol != htons(ETH_P_IP));
1744	BUG_ON(ih->protocol != IPPROTO_TCP);
1745	BUG_ON(skb->ip_summed != CHECKSUM_PARTIAL);
1746	BUG_ON(num_frags > LEPP_MAX_FRAGS);
1747	/*--BUG_ON(num_segs != (d_len + (p_len - 1)) / p_len); */
1748	BUG_ON(num_segs <= 1);
1749
1750
1751	/* Finish preparing the command. */
1752
1753	/* Copy the command header. */
1754	*cmd = cmd_init;
1755
1756	/* Copy the "header". */
1757	memcpy(&cmd->frags[num_frags], data, sh_len);
1758
1759
1760	/* Prefetch and wait, to minimize time spent holding the spinlock. */
1761	prefetch_L1(&eq->comp_tail);
1762	prefetch_L1(&eq->cmd_tail);
1763	mb();
1764
1765
1766	/* Enqueue the command. */
1767
1768	spin_lock_irqsave(&priv->eq_lock, irqflags);
1769
1770	/* Handle completions if needed to make room. */
1771	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
1772	if (lepp_num_free_comp_slots(eq) == 0) {
1773		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
1774		if (nolds == 0) {
1775busy:
1776			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1777			return NETDEV_TX_BUSY;
1778		}
1779	}
1780
1781	cmd_head = eq->cmd_head;
1782	cmd_tail = eq->cmd_tail;
1783
1784	/* Prepare to advance, detecting full queue. */
1785	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
1786	cmd_next = cmd_tail + cmd_size;
1787	if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1788		goto busy;
1789	if (cmd_next > LEPP_CMD_LIMIT) {
1790		cmd_next = 0;
1791		if (cmd_next == cmd_head)
1792			goto busy;
1793	}
1794
1795	/* Copy the command. */
1796	memcpy(&eq->cmds[cmd_tail], cmd, cmd_size);
1797
1798	/* Advance. */
1799	cmd_tail = cmd_next;
1800
1801	/* Record "skb" for eventual freeing. */
1802	comp_tail = eq->comp_tail;
1803	eq->comps[comp_tail] = skb;
1804	LEPP_QINC(comp_tail);
1805	eq->comp_tail = comp_tail;
1806
1807	/* Flush before allowing LEPP to handle the command. */
1808	/* ISSUE: Is this the optimal location for the flush? */
1809	__insn_mf();
1810
1811	eq->cmd_tail = cmd_tail;
1812
1813	/* NOTE: Using "4" here is more efficient than "0" or "2", */
1814	/* and, strangely, more efficient than pre-checking the number */
1815	/* of available completions, and comparing it to 4. */
1816	if (nolds == 0)
1817		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
1818
1819	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1820
1821	/* Handle completions. */
1822	for (i = 0; i < nolds; i++)
1823		dev_consume_skb_any(olds[i]);
1824
1825	/* Update stats. */
1826	u64_stats_update_begin(&stats->syncp);
1827	stats->tx_packets += num_segs;
1828	stats->tx_bytes += (num_segs * sh_len) + d_len;
1829	u64_stats_update_end(&stats->syncp);
1830
1831	/* Make sure the egress timer is scheduled. */
1832	tile_net_schedule_egress_timer(info);
1833
1834	return NETDEV_TX_OK;
1835}
1836
1837
1838/*
1839 * Transmit a packet (called by the kernel via "hard_start_xmit" hook).
1840 */
1841static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
1842{
1843	struct tile_net_priv *priv = netdev_priv(dev);
1844	int my_cpu = smp_processor_id();
1845	struct tile_net_cpu *info = priv->cpu[my_cpu];
1846	struct tile_net_stats_t *stats = &info->stats;
1847
1848	unsigned long irqflags;
1849
1850	struct skb_shared_info *sh = skb_shinfo(skb);
1851
1852	unsigned int len = skb->len;
1853	unsigned char *data = skb->data;
1854
1855	unsigned int csum_start = skb_checksum_start_offset(skb);
1856
1857	lepp_frag_t frags[1 + MAX_SKB_FRAGS];
1858
1859	unsigned int num_frags;
1860
1861	lepp_queue_t *eq = priv->eq;
1862
1863	struct sk_buff *olds[8];
1864	unsigned int wanted = 8;
1865	unsigned int i, nolds = 0;
1866
1867	unsigned int cmd_size = sizeof(lepp_cmd_t);
1868
1869	unsigned int cmd_head, cmd_tail, cmd_next;
1870	unsigned int comp_tail;
1871
1872	lepp_cmd_t cmds[1 + MAX_SKB_FRAGS];
1873
1874
1875	/*
1876	 * This is paranoia, since we think that if the link doesn't come
1877	 * up, telling Linux we have no carrier will keep it from trying
1878	 * to transmit.  If it does, though, we can't execute this routine,
1879	 * since data structures we depend on aren't set up yet.
1880	 */
1881	if (!info->registered)
1882		return NETDEV_TX_BUSY;
1883
1884
1885	/* Save the timestamp. */
1886	netif_trans_update(dev);
1887
1888
1889#ifdef TILE_NET_PARANOIA
1890#if CHIP_HAS_CBOX_HOME_MAP()
1891	if (hash_default) {
1892		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)data);
1893		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
1894			panic("Non-HFH egress buffer! VA=%p Mode=%d PTE=%llx",
1895			      data, hv_pte_get_mode(pte), hv_pte_val(pte));
1896	}
1897#endif
1898#endif
1899
1900
1901#ifdef TILE_NET_DUMP_PACKETS
1902	/* ISSUE: Does not dump the "frags". */
1903	dump_packet(data, skb_headlen(skb), "tx");
1904#endif /* TILE_NET_DUMP_PACKETS */
1905
1906
1907	if (sh->gso_size != 0)
1908		return tile_net_tx_tso(skb, dev);
1909
1910
1911	/* Prepare the commands. */
1912
1913	num_frags = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
1914
1915	for (i = 0; i < num_frags; i++) {
1916
1917		bool final = (i == num_frags - 1);
1918
1919		lepp_cmd_t cmd = {
1920			.cpa_lo = frags[i].cpa_lo,
1921			.cpa_hi = frags[i].cpa_hi,
1922			.length = frags[i].length,
1923			.hash_for_home = frags[i].hash_for_home,
1924			.send_completion = final,
1925			.end_of_packet = final
1926		};
1927
1928		if (i == 0 && skb->ip_summed == CHECKSUM_PARTIAL) {
1929			cmd.compute_checksum = 1;
1930			cmd.checksum_data.bits.start_byte = csum_start;
1931			cmd.checksum_data.bits.count = len - csum_start;
1932			cmd.checksum_data.bits.destination_byte =
1933				csum_start + skb->csum_offset;
1934		}
1935
1936		cmds[i] = cmd;
1937	}
1938
1939
1940	/* Prefetch and wait, to minimize time spent holding the spinlock. */
1941	prefetch_L1(&eq->comp_tail);
1942	prefetch_L1(&eq->cmd_tail);
1943	mb();
1944
1945
1946	/* Enqueue the commands. */
1947
1948	spin_lock_irqsave(&priv->eq_lock, irqflags);
1949
1950	/* Handle completions if needed to make room. */
1951	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
1952	if (lepp_num_free_comp_slots(eq) == 0) {
1953		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
1954		if (nolds == 0) {
1955busy:
1956			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
1957			return NETDEV_TX_BUSY;
1958		}
1959	}
1960
1961	cmd_head = eq->cmd_head;
1962	cmd_tail = eq->cmd_tail;
1963
1964	/* Copy the commands, or fail. */
1965	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
1966	for (i = 0; i < num_frags; i++) {
1967
1968		/* Prepare to advance, detecting full queue. */
1969		cmd_next = cmd_tail + cmd_size;
1970		if (cmd_tail < cmd_head && cmd_next >= cmd_head)
1971			goto busy;
1972		if (cmd_next > LEPP_CMD_LIMIT) {
1973			cmd_next = 0;
1974			if (cmd_next == cmd_head)
1975				goto busy;
1976		}
1977
1978		/* Copy the command. */
1979		*(lepp_cmd_t *)&eq->cmds[cmd_tail] = cmds[i];
1980
1981		/* Advance. */
1982		cmd_tail = cmd_next;
1983	}
1984
1985	/* Record "skb" for eventual freeing. */
1986	comp_tail = eq->comp_tail;
1987	eq->comps[comp_tail] = skb;
1988	LEPP_QINC(comp_tail);
1989	eq->comp_tail = comp_tail;
1990
1991	/* Flush before allowing LEPP to handle the command. */
1992	/* ISSUE: Is this the optimal location for the flush? */
1993	__insn_mf();
1994
1995	eq->cmd_tail = cmd_tail;
1996
1997	/* NOTE: Using "4" here is more efficient than "0" or "2", */
1998	/* and, strangely, more efficient than pre-checking the number */
1999	/* of available completions, and comparing it to 4. */
2000	if (nolds == 0)
2001		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
2002
2003	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
2004
2005	/* Handle completions. */
2006	for (i = 0; i < nolds; i++)
2007		dev_consume_skb_any(olds[i]);
2008
2009	/* HACK: Track "expanded" size for short packets (e.g. 42 < 60). */
2010	u64_stats_update_begin(&stats->syncp);
2011	stats->tx_packets++;
2012	stats->tx_bytes += ((len >= ETH_ZLEN) ? len : ETH_ZLEN);
2013	u64_stats_update_end(&stats->syncp);
2014
2015	/* Make sure the egress timer is scheduled. */
2016	tile_net_schedule_egress_timer(info);
2017
2018	return NETDEV_TX_OK;
2019}
2020
2021
2022/*
2023 * Deal with a transmit timeout.
2024 */
2025static void tile_net_tx_timeout(struct net_device *dev)
2026{
2027	PDEBUG("tile_net_tx_timeout()\n");
2028	PDEBUG("Transmit timeout at %ld, latency %ld\n", jiffies,
2029	       jiffies - dev_trans_start(dev));
2030
2031	/* XXX: ISSUE: This doesn't seem useful for us. */
2032	netif_wake_queue(dev);
2033}
2034
2035
2036/*
2037 * Ioctl commands.
2038 */
2039static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2040{
2041	return -EOPNOTSUPP;
2042}
2043
2044
2045/*
2046 * Get System Network Statistics.
2047 *
2048 * Returns the address of the device statistics structure.
2049 */
2050static struct rtnl_link_stats64 *tile_net_get_stats64(struct net_device *dev,
2051		struct rtnl_link_stats64 *stats)
2052{
2053	struct tile_net_priv *priv = netdev_priv(dev);
2054	u64 rx_packets = 0, tx_packets = 0;
2055	u64 rx_bytes = 0, tx_bytes = 0;
2056	u64 rx_errors = 0, rx_dropped = 0;
 
2057	int i;
2058
2059	for_each_online_cpu(i) {
2060		struct tile_net_stats_t *cpu_stats;
2061		u64 trx_packets, ttx_packets, trx_bytes, ttx_bytes;
2062		u64 trx_errors, trx_dropped;
2063		unsigned int start;
2064
2065		if (priv->cpu[i] == NULL)
2066			continue;
2067		cpu_stats = &priv->cpu[i]->stats;
2068
2069		do {
2070			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
2071			trx_packets = cpu_stats->rx_packets;
2072			ttx_packets = cpu_stats->tx_packets;
2073			trx_bytes   = cpu_stats->rx_bytes;
2074			ttx_bytes   = cpu_stats->tx_bytes;
2075			trx_errors  = cpu_stats->rx_errors;
2076			trx_dropped = cpu_stats->rx_dropped;
2077		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
2078
2079		rx_packets += trx_packets;
2080		tx_packets += ttx_packets;
2081		rx_bytes   += trx_bytes;
2082		tx_bytes   += ttx_bytes;
2083		rx_errors  += trx_errors;
2084		rx_dropped += trx_dropped;
2085	}
2086
2087	stats->rx_packets = rx_packets;
2088	stats->tx_packets = tx_packets;
2089	stats->rx_bytes   = rx_bytes;
2090	stats->tx_bytes   = tx_bytes;
2091	stats->rx_errors  = rx_errors;
2092	stats->rx_dropped = rx_dropped;
2093
2094	return stats;
2095}
2096
2097
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2098
2099/*
2100 * Change the Ethernet Address of the NIC.
2101 *
2102 * The hypervisor driver does not support changing MAC address.  However,
2103 * the IPP does not do anything with the MAC address, so the address which
2104 * gets used on outgoing packets, and which is accepted on incoming packets,
2105 * is completely up to the NetIO program or kernel driver which is actually
2106 * handling them.
2107 *
2108 * Returns 0 on success, negative on failure.
2109 */
2110static int tile_net_set_mac_address(struct net_device *dev, void *p)
2111{
2112	struct sockaddr *addr = p;
2113
2114	if (!is_valid_ether_addr(addr->sa_data))
2115		return -EADDRNOTAVAIL;
2116
2117	/* ISSUE: Note that "dev_addr" is now a pointer. */
2118	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
 
2119
2120	return 0;
2121}
2122
2123
2124/*
2125 * Obtain the MAC address from the hypervisor.
2126 * This must be done before opening the device.
2127 */
2128static int tile_net_get_mac(struct net_device *dev)
2129{
2130	struct tile_net_priv *priv = netdev_priv(dev);
2131
2132	char hv_dev_name[32];
2133	int len;
2134
2135	__netio_getset_offset_t offset = { .word = NETIO_IPP_PARAM_OFF };
2136
2137	int ret;
2138
2139	/* For example, "xgbe0". */
2140	strcpy(hv_dev_name, dev->name);
2141	len = strlen(hv_dev_name);
2142
2143	/* For example, "xgbe/0". */
2144	hv_dev_name[len] = hv_dev_name[len - 1];
2145	hv_dev_name[len - 1] = '/';
2146	len++;
2147
2148	/* For example, "xgbe/0/native_hash". */
2149	strcpy(hv_dev_name + len, hash_default ? "/native_hash" : "/native");
2150
2151	/* Get the hypervisor handle for this device. */
2152	priv->hv_devhdl = hv_dev_open((HV_VirtAddr)hv_dev_name, 0);
2153	PDEBUG("hv_dev_open(%s) returned %d %p\n",
2154	       hv_dev_name, priv->hv_devhdl, &priv->hv_devhdl);
2155	if (priv->hv_devhdl < 0) {
2156		if (priv->hv_devhdl == HV_ENODEV)
2157			printk(KERN_DEBUG "Ignoring unconfigured device %s\n",
2158				 hv_dev_name);
2159		else
2160			printk(KERN_DEBUG "hv_dev_open(%s) returned %d\n",
2161				 hv_dev_name, priv->hv_devhdl);
2162		return -1;
2163	}
2164
2165	/*
2166	 * Read the hardware address from the hypervisor.
2167	 * ISSUE: Note that "dev_addr" is now a pointer.
2168	 */
2169	offset.bits.class = NETIO_PARAM;
2170	offset.bits.addr = NETIO_PARAM_MAC;
2171	ret = hv_dev_pread(priv->hv_devhdl, 0,
2172			   (HV_VirtAddr)dev->dev_addr, dev->addr_len,
2173			   offset.word);
2174	PDEBUG("hv_dev_pread(NETIO_PARAM_MAC) returned %d\n", ret);
2175	if (ret <= 0) {
2176		printk(KERN_DEBUG "hv_dev_pread(NETIO_PARAM_MAC) %s failed\n",
2177		       dev->name);
2178		/*
2179		 * Since the device is configured by the hypervisor but we
2180		 * can't get its MAC address, we are most likely running
2181		 * the simulator, so let's generate a random MAC address.
2182		 */
2183		eth_hw_addr_random(dev);
2184	}
2185
2186	return 0;
2187}
2188
2189
2190#ifdef CONFIG_NET_POLL_CONTROLLER
2191/*
2192 * Polling 'interrupt' - used by things like netconsole to send skbs
2193 * without having to re-enable interrupts. It's not called while
2194 * the interrupt routine is executing.
2195 */
2196static void tile_net_netpoll(struct net_device *dev)
2197{
2198	struct tile_net_priv *priv = netdev_priv(dev);
2199	disable_percpu_irq(priv->intr_id);
2200	tile_net_handle_ingress_interrupt(priv->intr_id, dev);
2201	enable_percpu_irq(priv->intr_id, 0);
2202}
2203#endif
2204
2205
2206static const struct net_device_ops tile_net_ops = {
2207	.ndo_open = tile_net_open,
2208	.ndo_stop = tile_net_stop,
2209	.ndo_start_xmit = tile_net_tx,
2210	.ndo_do_ioctl = tile_net_ioctl,
2211	.ndo_get_stats64 = tile_net_get_stats64,
 
2212	.ndo_tx_timeout = tile_net_tx_timeout,
2213	.ndo_set_mac_address = tile_net_set_mac_address,
2214#ifdef CONFIG_NET_POLL_CONTROLLER
2215	.ndo_poll_controller = tile_net_netpoll,
2216#endif
2217};
2218
2219
2220/*
2221 * The setup function.
2222 *
2223 * This uses ether_setup() to assign various fields in dev, including
2224 * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
2225 */
2226static void tile_net_setup(struct net_device *dev)
2227{
2228	netdev_features_t features = 0;
2229
2230	ether_setup(dev);
 
2231	dev->netdev_ops = &tile_net_ops;
 
2232	dev->watchdog_timeo = TILE_NET_TIMEOUT;
2233	dev->tx_queue_len = TILE_NET_TX_QUEUE_LEN;
2234
2235	/* MTU range: 68 - 1500 */
2236	dev->mtu = TILE_NET_MTU;
2237	dev->min_mtu = ETH_MIN_MTU;
2238	dev->max_mtu = TILE_NET_MTU;
2239
2240	features |= NETIF_F_HW_CSUM;
2241	features |= NETIF_F_SG;
2242
2243	/* We support TSO iff the HV supports sufficient frags. */
2244	if (LEPP_MAX_FRAGS >= 1 + MAX_SKB_FRAGS)
2245		features |= NETIF_F_TSO;
 
 
 
 
 
 
 
2246
2247	/* We can't support HIGHDMA without hash_default, since we need
2248	 * to be able to finv() with a VA if we don't have hash_default.
2249	 */
2250	if (hash_default)
2251		features |= NETIF_F_HIGHDMA;
2252
2253	dev->hw_features   |= features;
2254	dev->vlan_features |= features;
2255	dev->features      |= features;
 
 
2256}
2257
2258
2259/*
2260 * Allocate the device structure, register the device, and obtain the
2261 * MAC address from the hypervisor.
2262 */
2263static struct net_device *tile_net_dev_init(const char *name)
2264{
2265	int ret;
2266	struct net_device *dev;
2267	struct tile_net_priv *priv;
2268
2269	/*
2270	 * Allocate the device structure.  This allocates "priv", calls
2271	 * tile_net_setup(), and saves "name".  Normally, "name" is a
2272	 * template, instantiated by register_netdev(), but not for us.
2273	 */
2274	dev = alloc_netdev(sizeof(*priv), name, NET_NAME_UNKNOWN,
2275			   tile_net_setup);
2276	if (!dev) {
2277		pr_err("alloc_netdev(%s) failed\n", name);
2278		return NULL;
2279	}
2280
2281	priv = netdev_priv(dev);
2282
2283	/* Initialize "priv". */
2284
2285	memset(priv, 0, sizeof(*priv));
2286
2287	/* Save "dev" for "tile_net_open_retry()". */
2288	priv->dev = dev;
2289
2290	INIT_DELAYED_WORK(&priv->retry_work, tile_net_open_retry);
2291
2292	spin_lock_init(&priv->eq_lock);
2293
2294	/* Allocate "eq". */
2295	priv->eq_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, EQ_ORDER);
2296	if (!priv->eq_pages) {
2297		free_netdev(dev);
2298		return NULL;
2299	}
2300	priv->eq = page_address(priv->eq_pages);
2301
2302	/* Register the network device. */
2303	ret = register_netdev(dev);
2304	if (ret) {
2305		pr_err("register_netdev %s failed %d\n", dev->name, ret);
2306		__free_pages(priv->eq_pages, EQ_ORDER);
2307		free_netdev(dev);
2308		return NULL;
2309	}
2310
2311	/* Get the MAC address. */
2312	ret = tile_net_get_mac(dev);
2313	if (ret < 0) {
2314		unregister_netdev(dev);
2315		__free_pages(priv->eq_pages, EQ_ORDER);
2316		free_netdev(dev);
2317		return NULL;
2318	}
2319
2320	return dev;
2321}
2322
2323
2324/*
2325 * Module cleanup.
2326 *
2327 * FIXME: If compiled as a module, this module cannot be "unloaded",
2328 * because the "ingress interrupt handler" is registered permanently.
2329 */
2330static void tile_net_cleanup(void)
2331{
2332	int i;
2333
2334	for (i = 0; i < TILE_NET_DEVS; i++) {
2335		if (tile_net_devs[i]) {
2336			struct net_device *dev = tile_net_devs[i];
2337			struct tile_net_priv *priv = netdev_priv(dev);
2338			unregister_netdev(dev);
2339			finv_buffer_remote(priv->eq, EQ_SIZE, 0);
2340			__free_pages(priv->eq_pages, EQ_ORDER);
2341			free_netdev(dev);
2342		}
2343	}
2344}
2345
2346
2347/*
2348 * Module initialization.
2349 */
2350static int tile_net_init_module(void)
2351{
2352	pr_info("Tilera Network Driver\n");
2353
2354	tile_net_devs[0] = tile_net_dev_init("xgbe0");
2355	tile_net_devs[1] = tile_net_dev_init("xgbe1");
2356	tile_net_devs[2] = tile_net_dev_init("gbe0");
2357	tile_net_devs[3] = tile_net_dev_init("gbe1");
2358
2359	return 0;
2360}
2361
2362
2363module_init(tile_net_init_module);
2364module_exit(tile_net_cleanup);
2365
2366
2367#ifndef MODULE
2368
2369/*
2370 * The "network_cpus" boot argument specifies the cpus that are dedicated
2371 * to handle ingress packets.
2372 *
2373 * The parameter should be in the form "network_cpus=m-n[,x-y]", where
2374 * m, n, x, y are integer numbers that represent the cpus that can be
2375 * neither a dedicated cpu nor a dataplane cpu.
2376 */
2377static int __init network_cpus_setup(char *str)
2378{
2379	int rc = cpulist_parse_crop(str, &network_cpus_map);
2380	if (rc != 0) {
2381		pr_warn("network_cpus=%s: malformed cpu list\n", str);
 
2382	} else {
2383
2384		/* Remove dedicated cpus. */
2385		cpumask_and(&network_cpus_map, &network_cpus_map,
2386			    cpu_possible_mask);
2387
2388
2389		if (cpumask_empty(&network_cpus_map)) {
2390			pr_warn("Ignoring network_cpus='%s'\n", str);
 
2391		} else {
2392			pr_info("Linux network CPUs: %*pbl\n",
2393				cpumask_pr_args(&network_cpus_map));
 
2394			network_cpus_used = true;
2395		}
2396	}
2397
2398	return 0;
2399}
2400__setup("network_cpus=", network_cpus_setup);
2401
2402#endif