tile: Move the Tilera driver

Move the Tilera driver into drivers/net/ethernet/tile and
make the necessary Kconfig and Makefile changes.

Updated the Kconfig so that the options defualt to y if TILE kernel.

CC: Chris Metcalf <cmetcalf@tilera.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

3 files changed, 2490 insertions, 0 deletions

diff --git a/drivers/net/ethernet/tile/Kconfig b/drivers/net/ethernet/tile/Kconfig
new file mode 100644
index 000000000000..2d9218f86bca
--- /dev/null
+++ b/drivers/net/ethernet/tile/Kconfig
@@ -0,0 +1,15 @@
+#
+# Tilera network device configuration
+#
+
+config TILE_NET
+	tristate "Tilera GBE/XGBE network driver support"
+	depends on TILE
+	default y
+	select CRC32
+	---help---
+	  This is a standard Linux network device driver for the
+	  on-chip Tilera Gigabit Ethernet and XAUI interfaces.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called tile_net.
diff --git a/drivers/net/ethernet/tile/Makefile b/drivers/net/ethernet/tile/Makefile
new file mode 100644
index 000000000000..f634f142cab4
--- /dev/null
+++ b/drivers/net/ethernet/tile/Makefile
@@ -0,0 +1,10 @@
+#
+# Makefile for the TILE on-chip networking support.
+#
+
+obj-$(CONFIG_TILE_NET) += tile_net.o
+ifdef CONFIG_TILEGX
+tile_net-objs := tilegx.o mpipe.o iorpc_mpipe.o dma_queue.o
+else
+tile_net-objs := tilepro.o
+endif
diff --git a/drivers/net/ethernet/tile/tilepro.c b/drivers/net/ethernet/tile/tilepro.c
new file mode 100644
index 000000000000..1e2af96fc29c
--- /dev/null
+++ b/drivers/net/ethernet/tile/tilepro.c
@@ -0,0 +1,2465 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/moduleparam.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>      /* printk() */
+#include <linux/slab.h>        /* kmalloc() */
+#include <linux/errno.h>       /* error codes */
+#include <linux/types.h>       /* size_t */
+#include <linux/interrupt.h>
+#include <linux/in.h>
+#include <linux/netdevice.h>   /* struct device, and other headers */
+#include <linux/etherdevice.h> /* eth_type_trans */
+#include <linux/skbuff.h>
+#include <linux/ioctl.h>
+#include <linux/cdev.h>
+#include <linux/hugetlb.h>
+#include <linux/in6.h>
+#include <linux/timer.h>
+#include <linux/io.h>
+#include <asm/checksum.h>
+#include <asm/homecache.h>
+
+#include <hv/drv_xgbe_intf.h>
+#include <hv/drv_xgbe_impl.h>
+#include <hv/hypervisor.h>
+#include <hv/netio_intf.h>
+
+/* For TSO */
+#include <linux/ip.h>
+#include <linux/tcp.h>
+
+
+/*
+ * First, "tile_net_init_module()" initializes all four "devices" which
+ * can be used by linux.
+ *
+ * Then, "ifconfig DEVICE up" calls "tile_net_open()", which analyzes
+ * the network cpus, then uses "tile_net_open_aux()" to initialize
+ * LIPP/LEPP, and then uses "tile_net_open_inner()" to register all
+ * the tiles, provide buffers to LIPP, allow ingress to start, and
+ * turn on hypervisor interrupt handling (and NAPI) on all tiles.
+ *
+ * If registration fails due to the link being down, then "retry_work"
+ * is used to keep calling "tile_net_open_inner()" until it succeeds.
+ *
+ * If "ifconfig DEVICE down" is called, it uses "tile_net_stop()" to
+ * stop egress, drain the LIPP buffers, unregister all the tiles, stop
+ * LIPP/LEPP, and wipe the LEPP queue.
+ *
+ * We start out with the ingress interrupt enabled on each CPU.  When
+ * this interrupt fires, we disable it, and call "napi_schedule()".
+ * This will cause "tile_net_poll()" to be called, which will pull
+ * packets from the netio queue, filtering them out, or passing them
+ * to "netif_receive_skb()".  If our budget is exhausted, we will
+ * return, knowing we will be called again later.  Otherwise, we
+ * reenable the ingress interrupt, and call "napi_complete()".
+ *
+ * HACK: Since disabling the ingress interrupt is not reliable, we
+ * ignore the interrupt if the global "active" flag is false.
+ *
+ *
+ * NOTE: The use of "native_driver" ensures that EPP exists, and that
+ * we are using "LIPP" and "LEPP".
+ *
+ * NOTE: Failing to free completions for an arbitrarily long time
+ * (which is defined to be illegal) does in fact cause bizarre
+ * problems.  The "egress_timer" helps prevent this from happening.
+ */
+
+
+/* HACK: Allow use of "jumbo" packets. */
+/* This should be 1500 if "jumbo" is not set in LIPP. */
+/* This should be at most 10226 (10240 - 14) if "jumbo" is set in LIPP. */
+/* ISSUE: This has not been thoroughly tested (except at 1500). */
+#define TILE_NET_MTU 1500
+
+/* HACK: Define to support GSO. */
+/* ISSUE: This may actually hurt performance of the TCP blaster. */
+/* #define TILE_NET_GSO */
+
+/* Define this to collapse "duplicate" acks. */
+/* #define IGNORE_DUP_ACKS */
+
+/* HACK: Define this to verify incoming packets. */
+/* #define TILE_NET_VERIFY_INGRESS */
+
+/* Use 3000 to enable the Linux Traffic Control (QoS) layer, else 0. */
+#define TILE_NET_TX_QUEUE_LEN 0
+
+/* Define to dump packets (prints out the whole packet on tx and rx). */
+/* #define TILE_NET_DUMP_PACKETS */
+
+/* Define to enable debug spew (all PDEBUG's are enabled). */
+/* #define TILE_NET_DEBUG */
+
+
+/* Define to activate paranoia checks. */
+/* #define TILE_NET_PARANOIA */
+
+/* Default transmit lockup timeout period, in jiffies. */
+#define TILE_NET_TIMEOUT (5 * HZ)
+
+/* Default retry interval for bringing up the NetIO interface, in jiffies. */
+#define TILE_NET_RETRY_INTERVAL (5 * HZ)
+
+/* Number of ports (xgbe0, xgbe1, gbe0, gbe1). */
+#define TILE_NET_DEVS 4
+
+
+
+/* Paranoia. */
+#if NET_IP_ALIGN != LIPP_PACKET_PADDING
+#error "NET_IP_ALIGN must match LIPP_PACKET_PADDING."
+#endif
+
+
+/* Debug print. */
+#ifdef TILE_NET_DEBUG
+#define PDEBUG(fmt, args...) net_printk(fmt, ## args)
+#else
+#define PDEBUG(fmt, args...)
+#endif
+
+
+MODULE_AUTHOR("Tilera");
+MODULE_LICENSE("GPL");
+
+
+/*
+ * Queue of incoming packets for a specific cpu and device.
+ *
+ * Includes a pointer to the "system" data, and the actual "user" data.
+ */
+struct tile_netio_queue {
+	netio_queue_impl_t *__system_part;
+	netio_queue_user_impl_t __user_part;
+
+};
+
+
+/*
+ * Statistics counters for a specific cpu and device.
+ */
+struct tile_net_stats_t {
+	u32 rx_packets;
+	u32 rx_bytes;
+	u32 tx_packets;
+	u32 tx_bytes;
+};
+
+
+/*
+ * Info for a specific cpu and device.
+ *
+ * ISSUE: There is a "dev" pointer in "napi" as well.
+ */
+struct tile_net_cpu {
+	/* The NAPI struct. */
+	struct napi_struct napi;
+	/* Packet queue. */
+	struct tile_netio_queue queue;
+	/* Statistics. */
+	struct tile_net_stats_t stats;
+	/* True iff NAPI is enabled. */
+	bool napi_enabled;
+	/* True if this tile has succcessfully registered with the IPP. */
+	bool registered;
+	/* True if the link was down last time we tried to register. */
+	bool link_down;
+	/* True if "egress_timer" is scheduled. */
+	bool egress_timer_scheduled;
+	/* Number of small sk_buffs which must still be provided. */
+	unsigned int num_needed_small_buffers;
+	/* Number of large sk_buffs which must still be provided. */
+	unsigned int num_needed_large_buffers;
+	/* A timer for handling egress completions. */
+	struct timer_list egress_timer;
+};
+
+
+/*
+ * Info for a specific device.
+ */
+struct tile_net_priv {
+	/* Our network device. */
+	struct net_device *dev;
+	/* Pages making up the egress queue. */
+	struct page *eq_pages;
+	/* Address of the actual egress queue. */
+	lepp_queue_t *eq;
+	/* Protects "eq". */
+	spinlock_t eq_lock;
+	/* The hypervisor handle for this interface. */
+	int hv_devhdl;
+	/* The intr bit mask that IDs this device. */
+	u32 intr_id;
+	/* True iff "tile_net_open_aux()" has succeeded. */
+	bool partly_opened;
+	/* True iff the device is "active". */
+	bool active;
+	/* Effective network cpus. */
+	struct cpumask network_cpus_map;
+	/* Number of network cpus. */
+	int network_cpus_count;
+	/* Credits per network cpu. */
+	int network_cpus_credits;
+	/* Network stats. */
+	struct net_device_stats stats;
+	/* For NetIO bringup retries. */
+	struct delayed_work retry_work;
+	/* Quick access to per cpu data. */
+	struct tile_net_cpu *cpu[NR_CPUS];
+};
+
+/* Log2 of the number of small pages needed for the egress queue. */
+#define EQ_ORDER  get_order(sizeof(lepp_queue_t))
+/* Size of the egress queue's pages. */
+#define EQ_SIZE   (1 << (PAGE_SHIFT + EQ_ORDER))
+
+/*
+ * The actual devices (xgbe0, xgbe1, gbe0, gbe1).
+ */
+static struct net_device *tile_net_devs[TILE_NET_DEVS];
+
+/*
+ * The "tile_net_cpu" structures for each device.
+ */
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe0);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe1);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe0);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe1);
+
+
+/*
+ * True if "network_cpus" was specified.
+ */
+static bool network_cpus_used;
+
+/*
+ * The actual cpus in "network_cpus".
+ */
+static struct cpumask network_cpus_map;
+
+
+
+#ifdef TILE_NET_DEBUG
+/*
+ * printk with extra stuff.
+ *
+ * We print the CPU we're running in brackets.
+ */
+static void net_printk(char *fmt, ...)
+{
+	int i;
+	int len;
+	va_list args;
+	static char buf[256];
+
+	len = sprintf(buf, "tile_net[%2.2d]: ", smp_processor_id());
+	va_start(args, fmt);
+	i = vscnprintf(buf + len, sizeof(buf) - len - 1, fmt, args);
+	va_end(args);
+	buf[255] = '\0';
+	pr_notice(buf);
+}
+#endif
+
+
+#ifdef TILE_NET_DUMP_PACKETS
+/*
+ * Dump a packet.
+ */
+static void dump_packet(unsigned char *data, unsigned long length, char *s)
+{
+	int my_cpu = smp_processor_id();
+
+	unsigned long i;
+	char buf[128];
+
+	static unsigned int count;
+
+	pr_info("dump_packet(data %p, length 0x%lx s %s count 0x%x)\n",
+	       data, length, s, count++);
+
+	pr_info("\n");
+
+	for (i = 0; i < length; i++) {
+		if ((i & 0xf) == 0)
+			sprintf(buf, "[%02d] %8.8lx:", my_cpu, i);
+		sprintf(buf + strlen(buf), " %2.2x", data[i]);
+		if ((i & 0xf) == 0xf || i == length - 1) {
+			strcat(buf, "\n");
+			pr_info("%s", buf);
+		}
+	}
+}
+#endif
+
+
+/*
+ * Provide support for the __netio_fastio1() swint
+ * (see <hv/drv_xgbe_intf.h> for how it is used).
+ *
+ * The fastio swint2 call may clobber all the caller-saved registers.
+ * It rarely clobbers memory, but we allow for the possibility in
+ * the signature just to be on the safe side.
+ *
+ * Also, gcc doesn't seem to allow an input operand to be
+ * clobbered, so we fake it with dummy outputs.
+ *
+ * This function can't be static because of the way it is declared
+ * in the netio header.
+ */
+inline int __netio_fastio1(u32 fastio_index, u32 arg0)
+{
+	long result, clobber_r1, clobber_r10;
+	asm volatile("swint2"
+		     : "=R00" (result),
+		       "=R01" (clobber_r1), "=R10" (clobber_r10)
+		     : "R10" (fastio_index), "R01" (arg0)
+		     : "memory", "r2", "r3", "r4",
+		       "r5", "r6", "r7", "r8", "r9",
+		       "r11", "r12", "r13", "r14",
+		       "r15", "r16", "r17", "r18", "r19",
+		       "r20", "r21", "r22", "r23", "r24",
+		       "r25", "r26", "r27", "r28", "r29");
+	return result;
+}
+
+
+/*
+ * Provide a linux buffer to LIPP.
+ */
+static void tile_net_provide_linux_buffer(struct tile_net_cpu *info,
+					  void *va, bool small)
+{
+	struct tile_netio_queue *queue = &info->queue;
+
+	/* Convert "va" and "small" to "linux_buffer_t". */
+	unsigned int buffer = ((unsigned int)(__pa(va) >> 7) << 1) + small;
+
+	__netio_fastio_free_buffer(queue->__user_part.__fastio_index, buffer);
+}
+
+
+/*
+ * Provide a linux buffer for LIPP.
+ *
+ * Note that the ACTUAL allocation for each buffer is a "struct sk_buff",
+ * plus a chunk of memory that includes not only the requested bytes, but
+ * also NET_SKB_PAD bytes of initial padding, and a "struct skb_shared_info".
+ *
+ * Note that "struct skb_shared_info" is 88 bytes with 64K pages and
+ * 268 bytes with 4K pages (since the frags[] array needs 18 entries).
+ *
+ * Without jumbo packets, the maximum packet size will be 1536 bytes,
+ * and we use 2 bytes (NET_IP_ALIGN) of padding.  ISSUE: If we told
+ * the hardware to clip at 1518 bytes instead of 1536 bytes, then we
+ * could save an entire cache line, but in practice, we don't need it.
+ *
+ * Since CPAs are 38 bits, and we can only encode the high 31 bits in
+ * a "linux_buffer_t", the low 7 bits must be zero, and thus, we must
+ * align the actual "va" mod 128.
+ *
+ * We assume that the underlying "head" will be aligned mod 64.  Note
+ * that in practice, we have seen "head" NOT aligned mod 128 even when
+ * using 2048 byte allocations, which is surprising.
+ *
+ * If "head" WAS always aligned mod 128, we could change LIPP to
+ * assume that the low SIX bits are zero, and the 7th bit is one, that
+ * is, align the actual "va" mod 128 plus 64, which would be "free".
+ *
+ * For now, the actual "head" pointer points at NET_SKB_PAD bytes of
+ * padding, plus 28 or 92 bytes of extra padding, plus the sk_buff
+ * pointer, plus the NET_IP_ALIGN padding, plus 126 or 1536 bytes for
+ * the actual packet, plus 62 bytes of empty padding, plus some
+ * padding and the "struct skb_shared_info".
+ *
+ * With 64K pages, a large buffer thus needs 32+92+4+2+1536+62+88
+ * bytes, or 1816 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 64K pages, a small buffer thus needs 32+92+4+2+126+88
+ * bytes, or 344 bytes, which means we are wasting 64+ bytes, and
+ * could presumably increase the size of small buffers.
+ *
+ * With 4K pages, a large buffer thus needs 32+92+4+2+1536+62+268
+ * bytes, or 1996 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 4K pages, a small buffer thus needs 32+92+4+2+126+268
+ * bytes, or 524 bytes, which is annoyingly wasteful.
+ *
+ * Maybe we should increase LIPP_SMALL_PACKET_SIZE to 192?
+ *
+ * ISSUE: Maybe we should increase "NET_SKB_PAD" to 64?
+ */
+static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
+					   bool small)
+{
+#if TILE_NET_MTU <= 1536
+	/* Without "jumbo", 2 + 1536 should be sufficient. */
+	unsigned int large_size = NET_IP_ALIGN + 1536;
+#else
+	/* ISSUE: This has not been tested. */
+	unsigned int large_size = NET_IP_ALIGN + TILE_NET_MTU + 100;
+#endif
+
+	/* Avoid "false sharing" with last cache line. */
+	/* ISSUE: This is already done by "dev_alloc_skb()". */
+	unsigned int len =
+		 (((small ? LIPP_SMALL_PACKET_SIZE : large_size) +
+		   CHIP_L2_LINE_SIZE() - 1) & -CHIP_L2_LINE_SIZE());
+
+	unsigned int padding = 128 - NET_SKB_PAD;
+	unsigned int align;
+
+	struct sk_buff *skb;
+	void *va;
+
+	struct sk_buff **skb_ptr;
+
+	/* Request 96 extra bytes for alignment purposes. */
+	skb = dev_alloc_skb(len + padding);
+	if (skb == NULL)
+		return false;
+
+	/* Skip 32 or 96 bytes to align "data" mod 128. */
+	align = -(long)skb->data & (128 - 1);
+	BUG_ON(align > padding);
+	skb_reserve(skb, align);
+
+	/* This address is given to IPP. */
+	va = skb->data;
+
+	/* Buffers must not span a huge page. */
+	BUG_ON(((((long)va & ~HPAGE_MASK) + len) & HPAGE_MASK) != 0);
+
+#ifdef TILE_NET_PARANOIA
+#if CHIP_HAS_CBOX_HOME_MAP()
+	if (hash_default) {
+		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
+		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
+			panic("Non-HFH ingress buffer! VA=%p Mode=%d PTE=%llx",
+			      va, hv_pte_get_mode(pte), hv_pte_val(pte));
+	}
+#endif
+#endif
+
+	/* Invalidate the packet buffer. */
+	if (!hash_default)
+		__inv_buffer(va, len);
+
+	/* Skip two bytes to satisfy LIPP assumptions. */
+	/* Note that this aligns IP on a 16 byte boundary. */
+	/* ISSUE: Do this when the packet arrives? */
+	skb_reserve(skb, NET_IP_ALIGN);
+
+	/* Save a back-pointer to 'skb'. */
+	skb_ptr = va - sizeof(*skb_ptr);
+	*skb_ptr = skb;
+
+	/* Make sure "skb_ptr" has been flushed. */
+	__insn_mf();
+
+	/* Provide the new buffer. */
+	tile_net_provide_linux_buffer(info, va, small);
+
+	return true;
+}
+
+
+/*
+ * Provide linux buffers for LIPP.
+ */
+static void tile_net_provide_needed_buffers(struct tile_net_cpu *info)
+{
+	while (info->num_needed_small_buffers != 0) {
+		if (!tile_net_provide_needed_buffer(info, true))
+			goto oops;
+		info->num_needed_small_buffers--;
+	}
+
+	while (info->num_needed_large_buffers != 0) {
+		if (!tile_net_provide_needed_buffer(info, false))
+			goto oops;
+		info->num_needed_large_buffers--;
+	}
+
+	return;
+
+oops:
+
+	/* Add a description to the page allocation failure dump. */
+	pr_notice("Could not provide a linux buffer to LIPP.\n");
+}
+
+
+/*
+ * Grab some LEPP completions, and store them in "comps", of size
+ * "comps_size", and return the number of completions which were
+ * stored, so the caller can free them.
+ */
+static unsigned int tile_net_lepp_grab_comps(lepp_queue_t *eq,
+					     struct sk_buff *comps[],
+					     unsigned int comps_size,
+					     unsigned int min_size)
+{
+	unsigned int n = 0;
+
+	unsigned int comp_head = eq->comp_head;
+	unsigned int comp_busy = eq->comp_busy;
+
+	while (comp_head != comp_busy && n < comps_size) {
+		comps[n++] = eq->comps[comp_head];
+		LEPP_QINC(comp_head);
+	}
+
+	if (n < min_size)
+		return 0;
+
+	eq->comp_head = comp_head;
+
+	return n;
+}
+
+
+/*
+ * Free some comps, and return true iff there are still some pending.
+ */
+static bool tile_net_lepp_free_comps(struct net_device *dev, bool all)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	lepp_queue_t *eq = priv->eq;
+
+	struct sk_buff *olds[64];
+	unsigned int wanted = 64;
+	unsigned int i, n;
+	bool pending;
+
+	spin_lock(&priv->eq_lock);
+
+	if (all)
+		eq->comp_busy = eq->comp_tail;
+
+	n = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+
+	pending = (eq->comp_head != eq->comp_tail);
+
+	spin_unlock(&priv->eq_lock);
+
+	for (i = 0; i < n; i++)
+		kfree_skb(olds[i]);
+
+	return pending;
+}
+
+
+/*
+ * Make sure the egress timer is scheduled.
+ *
+ * Note that we use "schedule if not scheduled" logic instead of the more
+ * obvious "reschedule" logic, because "reschedule" is fairly expensive.
+ */
+static void tile_net_schedule_egress_timer(struct tile_net_cpu *info)
+{
+	if (!info->egress_timer_scheduled) {
+		mod_timer_pinned(&info->egress_timer, jiffies + 1);
+		info->egress_timer_scheduled = true;
+	}
+}
+
+
+/*
+ * The "function" for "info->egress_timer".
+ *
+ * This timer will reschedule itself as long as there are any pending
+ * completions expected (on behalf of any tile).
+ *
+ * ISSUE: Realistically, will the timer ever stop scheduling itself?
+ *
+ * ISSUE: This timer is almost never actually needed, so just use a global
+ * timer that can run on any tile.
+ *
+ * ISSUE: Maybe instead track number of expected completions, and free
+ * only that many, resetting to zero if "pending" is ever false.
+ */
+static void tile_net_handle_egress_timer(unsigned long arg)
+{
+	struct tile_net_cpu *info = (struct tile_net_cpu *)arg;
+	struct net_device *dev = info->napi.dev;
+
+	/* The timer is no longer scheduled. */
+	info->egress_timer_scheduled = false;
+
+	/* Free comps, and reschedule timer if more are pending. */
+	if (tile_net_lepp_free_comps(dev, false))
+		tile_net_schedule_egress_timer(info);
+}
+
+
+#ifdef IGNORE_DUP_ACKS
+
+/*
+ * Help detect "duplicate" ACKs.  These are sequential packets (for a
+ * given flow) which are exactly 66 bytes long, sharing everything but
+ * ID=2@0x12, Hsum=2@0x18, Ack=4@0x2a, WinSize=2@0x30, Csum=2@0x32,
+ * Tstamps=10@0x38.  The ID's are +1, the Hsum's are -1, the Ack's are
+ * +N, and the Tstamps are usually identical.
+ *
+ * NOTE: Apparently truly duplicate acks (with identical "ack" values),
+ * should not be collapsed, as they are used for some kind of flow control.
+ */
+static bool is_dup_ack(char *s1, char *s2, unsigned int len)
+{
+	int i;
+
+	unsigned long long ignorable = 0;
+
+	/* Identification. */
+	ignorable |= (1ULL << 0x12);
+	ignorable |= (1ULL << 0x13);
+
+	/* Header checksum. */
+	ignorable |= (1ULL << 0x18);
+	ignorable |= (1ULL << 0x19);
+
+	/* ACK. */
+	ignorable |= (1ULL << 0x2a);
+	ignorable |= (1ULL << 0x2b);
+	ignorable |= (1ULL << 0x2c);
+	ignorable |= (1ULL << 0x2d);
+
+	/* WinSize. */
+	ignorable |= (1ULL << 0x30);
+	ignorable |= (1ULL << 0x31);
+
+	/* Checksum. */
+	ignorable |= (1ULL << 0x32);
+	ignorable |= (1ULL << 0x33);
+
+	for (i = 0; i < len; i++, ignorable >>= 1) {
+
+		if ((ignorable & 1) || (s1[i] == s2[i]))
+			continue;
+
+#ifdef TILE_NET_DEBUG
+		/* HACK: Mention non-timestamp diffs. */
+		if (i < 0x38 && i != 0x2f &&
+		    net_ratelimit())
+			pr_info("Diff at 0x%x\n", i);
+#endif
+
+		return false;
+	}
+
+#ifdef TILE_NET_NO_SUPPRESS_DUP_ACKS
+	/* HACK: Do not suppress truly duplicate ACKs. */
+	/* ISSUE: Is this actually necessary or helpful? */
+	if (s1[0x2a] == s2[0x2a] &&
+	    s1[0x2b] == s2[0x2b] &&
+	    s1[0x2c] == s2[0x2c] &&
+	    s1[0x2d] == s2[0x2d]) {
+		return false;
+	}
+#endif
+
+	return true;
+}
+
+#endif
+
+
+
+static void tile_net_discard_aux(struct tile_net_cpu *info, int index)
+{
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	int index2_aux = index + sizeof(netio_pkt_t);
+	int index2 =
+		((index2_aux ==
+		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
+		 0 : index2_aux);
+
+	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
+
+	/* Extract the "linux_buffer_t". */
+	unsigned int buffer = pkt->__packet.word;
+
+	/* Convert "linux_buffer_t" to "va". */
+	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+	/* Acquire the associated "skb". */
+	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+	struct sk_buff *skb = *skb_ptr;
+
+	kfree_skb(skb);
+
+	/* Consume this packet. */
+	qup->__packet_receive_read = index2;
+}
+
+
+/*
+ * Like "tile_net_poll()", but just discard packets.
+ */
+static void tile_net_discard_packets(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	while (qup->__packet_receive_read !=
+	       qsp->__packet_receive_queue.__packet_write) {
+		int index = qup->__packet_receive_read;
+		tile_net_discard_aux(info, index);
+	}
+}
+
+
+/*
+ * Handle the next packet.  Return true if "processed", false if "filtered".
+ */
+static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
+{
+	struct net_device *dev = info->napi.dev;
+
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+	struct tile_net_stats_t *stats = &info->stats;
+
+	int filter;
+
+	int index2_aux = index + sizeof(netio_pkt_t);
+	int index2 =
+		((index2_aux ==
+		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
+		 0 : index2_aux);
+
+	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
+
+	netio_pkt_metadata_t *metadata = NETIO_PKT_METADATA(pkt);
+
+	/* Extract the packet size.  FIXME: Shouldn't the second line */
+	/* get subtracted?  Mostly moot, since it should be "zero". */
+	unsigned long len =
+		(NETIO_PKT_CUSTOM_LENGTH(pkt) +
+		 NET_IP_ALIGN - NETIO_PACKET_PADDING);
+
+	/* Extract the "linux_buffer_t". */
+	unsigned int buffer = pkt->__packet.word;
+
+	/* Extract "small" (vs "large"). */
+	bool small = ((buffer & 1) != 0);
+
+	/* Convert "linux_buffer_t" to "va". */
+	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+	/* Extract the packet data pointer. */
+	/* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
+	unsigned char *buf = va + NET_IP_ALIGN;
+
+	/* Invalidate the packet buffer. */
+	if (!hash_default)
+		__inv_buffer(buf, len);
+
+	/* ISSUE: Is this needed? */
+	dev->last_rx = jiffies;
+
+#ifdef TILE_NET_DUMP_PACKETS
+	dump_packet(buf, len, "rx");
+#endif /* TILE_NET_DUMP_PACKETS */
+
+#ifdef TILE_NET_VERIFY_INGRESS
+	if (!NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt) &&
+	    NETIO_PKT_L4_CSUM_CALCULATED_M(metadata, pkt)) {
+		/* Bug 6624: Includes UDP packets with a "zero" checksum. */
+		pr_warning("Bad L4 checksum on %d byte packet.\n", len);
+	}
+	if (!NETIO_PKT_L3_CSUM_CORRECT_M(metadata, pkt) &&
+	    NETIO_PKT_L3_CSUM_CALCULATED_M(metadata, pkt)) {
+		dump_packet(buf, len, "rx");
+		panic("Bad L3 checksum.");
+	}
+	switch (NETIO_PKT_STATUS_M(metadata, pkt)) {
+	case NETIO_PKT_STATUS_OVERSIZE:
+		if (len >= 64) {
+			dump_packet(buf, len, "rx");
+			panic("Unexpected OVERSIZE.");
+		}
+		break;
+	case NETIO_PKT_STATUS_BAD:
+		pr_warning("Unexpected BAD %ld byte packet.\n", len);
+	}
+#endif
+
+	filter = 0;
+
+	/* ISSUE: Filter TCP packets with "bad" checksums? */
+
+	if (!(dev->flags & IFF_UP)) {
+		/* Filter packets received before we're up. */
+		filter = 1;
+	} else if (NETIO_PKT_STATUS_M(metadata, pkt) == NETIO_PKT_STATUS_BAD) {
+		/* Filter "truncated" packets. */
+		filter = 1;
+	} else if (!(dev->flags & IFF_PROMISC)) {
+		/* FIXME: Implement HW multicast filter. */
+		if (!is_multicast_ether_addr(buf)) {
+			/* Filter packets not for our address. */
+			const u8 *mine = dev->dev_addr;
+			filter = compare_ether_addr(mine, buf);
+		}
+	}
+
+	if (filter) {
+
+		/* ISSUE: Update "drop" statistics? */
+
+		tile_net_provide_linux_buffer(info, va, small);
+
+	} else {
+
+		/* Acquire the associated "skb". */
+		struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+		struct sk_buff *skb = *skb_ptr;
+
+		/* Paranoia. */
+		if (skb->data != buf)
+			panic("Corrupt linux buffer from LIPP! "
+			      "VA=%p, skb=%p, skb->data=%p\n",
+			      va, skb, skb->data);
+
+		/* Encode the actual packet length. */
+		skb_put(skb, len);
+
+		/* NOTE: This call also sets "skb->dev = dev". */
+		skb->protocol = eth_type_trans(skb, dev);
+
+		/* Avoid recomputing "good" TCP/UDP checksums. */
+		if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
+			skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+		netif_receive_skb(skb);
+
+		stats->rx_packets++;
+		stats->rx_bytes += len;
+
+		if (small)
+			info->num_needed_small_buffers++;
+		else
+			info->num_needed_large_buffers++;
+	}
+
+	/* Return four credits after every fourth packet. */
+	if (--qup->__receive_credit_remaining == 0) {
+		u32 interval = qup->__receive_credit_interval;
+		qup->__receive_credit_remaining = interval;
+		__netio_fastio_return_credits(qup->__fastio_index, interval);
+	}
+
+	/* Consume this packet. */
+	qup->__packet_receive_read = index2;
+
+	return !filter;
+}
+
+
+/*
+ * Handle some packets for the given device on the current CPU.
+ *
+ * If "tile_net_stop()" is called on some other tile while this
+ * function is running, we will return, hopefully before that
+ * other tile asks us to call "napi_disable()".
+ *
+ * The "rotting packet" race condition occurs if a packet arrives
+ * during the extremely narrow window between the queue appearing to
+ * be empty, and the ingress interrupt being re-enabled.  This happens
+ * a LOT under heavy network load.
+ */
+static int tile_net_poll(struct napi_struct *napi, int budget)
+{
+	struct net_device *dev = napi->dev;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	unsigned int work = 0;
+
+	while (priv->active) {
+		int index = qup->__packet_receive_read;
+		if (index == qsp->__packet_receive_queue.__packet_write)
+			break;
+
+		if (tile_net_poll_aux(info, index)) {
+			if (++work >= budget)
+				goto done;
+		}
+	}
+
+	napi_complete(&info->napi);
+
+	if (!priv->active)
+		goto done;
+
+	/* Re-enable the ingress interrupt. */
+	enable_percpu_irq(priv->intr_id);
+
+	/* HACK: Avoid the "rotting packet" problem (see above). */
+	if (qup->__packet_receive_read !=
+	    qsp->__packet_receive_queue.__packet_write) {
+		/* ISSUE: Sometimes this returns zero, presumably */
+		/* because an interrupt was handled for this tile. */
+		(void)napi_reschedule(&info->napi);
+	}
+
+done:
+
+	if (priv->active)
+		tile_net_provide_needed_buffers(info);
+
+	return work;
+}
+
+
+/*
+ * Handle an ingress interrupt for the given device on the current cpu.
+ *
+ * ISSUE: Sometimes this gets called after "disable_percpu_irq()" has
+ * been called!  This is probably due to "pending hypervisor downcalls".
+ *
+ * ISSUE: Is there any race condition between the "napi_schedule()" here
+ * and the "napi_complete()" call above?
+ */
+static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
+
+	/* Ignore unwanted interrupts. */
+	if (!priv->active)
+		return IRQ_HANDLED;
+
+	/* ISSUE: Sometimes "info->napi_enabled" is false here. */
+
+	napi_schedule(&info->napi);
+
+	return IRQ_HANDLED;
+}
+
+
+/*
+ * One time initialization per interface.
+ */
+static int tile_net_open_aux(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	int ret;
+	int dummy;
+	unsigned int epp_lotar;
+
+	/*
+	 * Find out where EPP memory should be homed.
+	 */
+	ret = hv_dev_pread(priv->hv_devhdl, 0,
+			   (HV_VirtAddr)&epp_lotar, sizeof(epp_lotar),
+			   NETIO_EPP_SHM_OFF);
+	if (ret < 0) {
+		pr_err("could not read epp_shm_queue lotar.\n");
+		return -EIO;
+	}
+
+	/*
+	 * Home the page on the EPP.
+	 */
+	{
+		int epp_home = hv_lotar_to_cpu(epp_lotar);
+		homecache_change_page_home(priv->eq_pages, EQ_ORDER, epp_home);
+	}
+
+	/*
+	 * Register the EPP shared memory queue.
+	 */
+	{
+		netio_ipp_address_t ea = {
+			.va = 0,
+			.pa = __pa(priv->eq),
+			.pte = hv_pte(0),
+			.size = EQ_SIZE,
+		};
+		ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
+		ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
+		ret = hv_dev_pwrite(priv->hv_devhdl, 0,
+				    (HV_VirtAddr)&ea,
+				    sizeof(ea),
+				    NETIO_EPP_SHM_OFF);
+		if (ret < 0)
+			return -EIO;
+	}
+
+	/*
+	 * Start LIPP/LEPP.
+	 */
+	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			  sizeof(dummy), NETIO_IPP_START_SHIM_OFF) < 0) {
+		pr_warning("Failed to start LIPP/LEPP.\n");
+		return -EIO;
+	}
+
+	return 0;
+}
+
+
+/*
+ * Register with hypervisor on the current CPU.
+ *
+ * Strangely, this function does important things even if it "fails",
+ * which is especially common if the link is not up yet.  Hopefully
+ * these things are all "harmless" if done twice!
+ */
+static void tile_net_register(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info;
+
+	struct tile_netio_queue *queue;
+
+	/* Only network cpus can receive packets. */
+	int queue_id =
+		cpumask_test_cpu(my_cpu, &priv->network_cpus_map) ? 0 : 255;
+
+	netio_input_config_t config = {
+		.flags = 0,
+		.num_receive_packets = priv->network_cpus_credits,
+		.queue_id = queue_id
+	};
+
+	int ret = 0;
+	netio_queue_impl_t *queuep;
+
+	PDEBUG("tile_net_register(queue_id %d)\n", queue_id);
+
+	if (!strcmp(dev->name, "xgbe0"))
+		info = &__get_cpu_var(hv_xgbe0);
+	else if (!strcmp(dev->name, "xgbe1"))
+		info = &__get_cpu_var(hv_xgbe1);
+	else if (!strcmp(dev->name, "gbe0"))
+		info = &__get_cpu_var(hv_gbe0);
+	else if (!strcmp(dev->name, "gbe1"))
+		info = &__get_cpu_var(hv_gbe1);
+	else
+		BUG();
+
+	/* Initialize the egress timer. */
+	init_timer(&info->egress_timer);
+	info->egress_timer.data = (long)info;
+	info->egress_t