Add the rt linux 4.1.3-rt3 as base

Import the rt linux 4.1.3-rt3 as OPNFV kvm base.

It's from git://git.kernel.org/pub/scm/linux/kernel/git/rt/linux-rt-devel.git linux-4.1.y-rt and
the base is:

commit 0917f823c59692d751951bf5ea699a2d1e2f26a2
Author: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date:   Sat Jul 25 12:13:34 2015 +0200

    Prepare v4.1.3-rt3

    Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

We lose all the git history this way and it's not good. We
should apply another opnfv project repo in future.

Change-Id: I87543d81c9df70d99c5001fbdf646b202c19f423
Signed-off-by: Yunhong Jiang <yunhong.jiang@intel.com>

42 files changed, 11160 insertions, 0 deletions

diff --git a/kernel/arch/ia64/sn/Makefile b/kernel/arch/ia64/sn/Makefile
new file mode 100644
index 000000000..79a7df02e
--- /dev/null
+++ b/kernel/arch/ia64/sn/Makefile
@@ -0,0 +1,12 @@
+# arch/ia64/sn/Makefile
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# Copyright (C) 2004 Silicon Graphics, Inc.  All Rights Reserved.
+#
+# Makefile for the sn ia64 subplatform
+#
+
+obj-y += kernel/ pci/
diff --git a/kernel/arch/ia64/sn/include/ioerror.h b/kernel/arch/ia64/sn/include/ioerror.h
new file mode 100644
index 000000000..e68f2b078
--- /dev/null
+++ b/kernel/arch/ia64/sn/include/ioerror.h
@@ -0,0 +1,81 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992 - 1997, 2000-2003 Silicon Graphics, Inc. All rights reserved.
+ */
+#ifndef _ASM_IA64_SN_IOERROR_H
+#define _ASM_IA64_SN_IOERROR_H
+
+/*
+ * IO error structure.
+ *
+ * This structure would expand to hold the information retrieved from
+ * all IO related error registers.
+ *
+ * This structure is defined to hold all system specific
+ * information related to a single error.
+ *
+ * This serves a couple of purpose.
+ *      - Error handling often involves translating one form of address to other
+ *        form. So, instead of having different data structures at each level,
+ *        we have a single structure, and the appropriate fields get filled in
+ *        at each layer.
+ *      - This provides a way to dump all error related information in any layer
+ *        of erorr handling (debugging aid).
+ *
+ * A second possibility is to allow each layer to define its own error
+ * data structure, and fill in the proper fields. This has the advantage
+ * of isolating the layers.
+ * A big concern is the potential stack usage (and overflow), if each layer
+ * defines these structures on stack (assuming we don't want to do kmalloc.
+ *
+ * Any layer wishing to pass extra information to a layer next to it in
+ * error handling hierarchy, can do so as a separate parameter.
+ */
+
+typedef struct io_error_s {
+    /* Bit fields indicating which structure fields are valid */
+    union {
+	struct {
+	    unsigned                ievb_errortype:1;
+	    unsigned                ievb_widgetnum:1;
+	    unsigned                ievb_widgetdev:1;
+	    unsigned                ievb_srccpu:1;
+	    unsigned                ievb_srcnode:1;
+	    unsigned                ievb_errnode:1;
+	    unsigned                ievb_sysioaddr:1;
+	    unsigned                ievb_xtalkaddr:1;
+	    unsigned                ievb_busspace:1;
+	    unsigned                ievb_busaddr:1;
+	    unsigned                ievb_vaddr:1;
+	    unsigned                ievb_memaddr:1;
+	    unsigned		    ievb_epc:1;
+	    unsigned		    ievb_ef:1;
+	    unsigned		    ievb_tnum:1;
+	} iev_b;
+	unsigned                iev_a;
+    } ie_v;
+
+    short                   ie_errortype;	/* error type: extra info about error */
+    short                   ie_widgetnum;	/* Widget number that's in error */
+    short                   ie_widgetdev;	/* Device within widget in error */
+    cpuid_t                 ie_srccpu;	/* CPU on srcnode generating error */
+    cnodeid_t               ie_srcnode;		/* Node which caused the error   */
+    cnodeid_t               ie_errnode;		/* Node where error was noticed  */
+    iopaddr_t               ie_sysioaddr;	/* Sys specific IO address       */
+    iopaddr_t               ie_xtalkaddr;	/* Xtalk (48bit) addr of Error   */
+    iopaddr_t               ie_busspace;	/* Bus specific address space    */
+    iopaddr_t               ie_busaddr;		/* Bus specific address          */
+    caddr_t                 ie_vaddr;	/* Virtual address of error      */
+    iopaddr_t               ie_memaddr;		/* Physical memory address       */
+    caddr_t		    ie_epc;		/* pc when error reported	 */
+    caddr_t		    ie_ef;		/* eframe when error reported	 */
+    short		    ie_tnum;		/* Xtalk TNUM field */
+} ioerror_t;
+
+#define	IOERROR_INIT(e)		do { (e)->ie_v.iev_a = 0; } while (0)
+#define	IOERROR_SETVALUE(e,f,v)	do { (e)->ie_ ## f = (v); (e)->ie_v.iev_b.ievb_ ## f = 1; } while (0)
+
+#endif /* _ASM_IA64_SN_IOERROR_H */
diff --git a/kernel/arch/ia64/sn/include/tio.h b/kernel/arch/ia64/sn/include/tio.h
new file mode 100644
index 000000000..6b2e7b75e
--- /dev/null
+++ b/kernel/arch/ia64/sn/include/tio.h
@@ -0,0 +1,41 @@
+/* 
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#ifndef _ASM_IA64_SN_TIO_H
+#define _ASM_IA64_SN_TIO_H
+
+#define	TIO_MMR_ADDR_MOD
+
+#define TIO_NODE_ID     TIO_MMR_ADDR_MOD(0x0000000090060e80)
+
+#define TIO_ITTE_BASE   0xb0008800        /* base of translation table entries */
+#define TIO_ITTE(bigwin)        (TIO_ITTE_BASE + 8*(bigwin))
+
+#define TIO_ITTE_OFFSET_BITS    8       /* size of offset field */
+#define TIO_ITTE_OFFSET_MASK    ((1<<TIO_ITTE_OFFSET_BITS)-1)
+#define TIO_ITTE_OFFSET_SHIFT   0
+
+#define TIO_ITTE_WIDGET_BITS    2       /* size of widget field */
+#define TIO_ITTE_WIDGET_MASK    ((1<<TIO_ITTE_WIDGET_BITS)-1)
+#define TIO_ITTE_WIDGET_SHIFT   12
+#define TIO_ITTE_VALID_MASK	0x1
+#define TIO_ITTE_VALID_SHIFT	16
+
+#define TIO_ITTE_WIDGET(itte) \
+	(((itte) >> TIO_ITTE_WIDGET_SHIFT) & TIO_ITTE_WIDGET_MASK)
+#define TIO_ITTE_VALID(itte) \
+	(((itte) >> TIO_ITTE_VALID_SHIFT) & TIO_ITTE_VALID_MASK)
+
+#define TIO_ITTE_PUT(nasid, bigwin, widget, addr, valid) \
+        REMOTE_HUB_S((nasid), TIO_ITTE(bigwin), \
+                (((((addr) >> TIO_BWIN_SIZE_BITS) & \
+                   TIO_ITTE_OFFSET_MASK) << TIO_ITTE_OFFSET_SHIFT) | \
+                (((widget) & TIO_ITTE_WIDGET_MASK) << TIO_ITTE_WIDGET_SHIFT)) | \
+		(( (valid) & TIO_ITTE_VALID_MASK) << TIO_ITTE_VALID_SHIFT))
+
+#endif /*  _ASM_IA64_SN_TIO_H */
diff --git a/kernel/arch/ia64/sn/include/xtalk/hubdev.h b/kernel/arch/ia64/sn/include/xtalk/hubdev.h
new file mode 100644
index 000000000..8182583c7
--- /dev/null
+++ b/kernel/arch/ia64/sn/include/xtalk/hubdev.h
@@ -0,0 +1,91 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992 - 1997, 2000-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+#ifndef _ASM_IA64_SN_XTALK_HUBDEV_H
+#define _ASM_IA64_SN_XTALK_HUBDEV_H
+
+#include "xtalk/xwidgetdev.h"
+
+#define HUB_WIDGET_ID_MAX 0xf
+#define DEV_PER_WIDGET (2*2*8)
+#define IIO_ITTE_WIDGET_BITS    4       /* size of widget field */
+#define IIO_ITTE_WIDGET_MASK    ((1<<IIO_ITTE_WIDGET_BITS)-1)
+#define IIO_ITTE_WIDGET_SHIFT   8
+
+#define IIO_ITTE_WIDGET(itte)	\
+	(((itte) >> IIO_ITTE_WIDGET_SHIFT) & IIO_ITTE_WIDGET_MASK)
+
+/*
+ * Use the top big window as a surrogate for the first small window
+ */
+#define SWIN0_BIGWIN            HUB_NUM_BIG_WINDOW
+#define IIO_NUM_ITTES   7
+#define HUB_NUM_BIG_WINDOW      (IIO_NUM_ITTES - 1)
+
+/* This struct is shared between the PROM and the kernel.
+ * Changes to this struct will require corresponding changes to the kernel.
+ */
+struct sn_flush_device_common {
+	int sfdl_bus;
+	int sfdl_slot;
+	int sfdl_pin;
+	struct common_bar_list {
+		unsigned long start;
+		unsigned long end;
+	} sfdl_bar_list[6];
+	unsigned long sfdl_force_int_addr;
+	unsigned long sfdl_flush_value;
+	volatile unsigned long *sfdl_flush_addr;
+	u32 sfdl_persistent_busnum;
+	u32 sfdl_persistent_segment;
+	struct pcibus_info *sfdl_pcibus_info;
+};
+
+/* This struct is kernel only and is not used by the PROM */
+struct sn_flush_device_kernel {
+	spinlock_t sfdl_flush_lock;
+	struct sn_flush_device_common *common;
+};
+
+/* 01/16/06 This struct is the old PROM/kernel struct and needs to be included
+ * for older official PROMs to function on the new kernel base.  This struct
+ * will be removed when the next official PROM release occurs. */
+
+struct sn_flush_device_war {
+	struct sn_flush_device_common common;
+	u32 filler; /* older PROMs expect the default size of a spinlock_t */
+};
+
+/*
+ * **widget_p - Used as an array[wid_num][device] of sn_flush_device_kernel.
+ */
+struct sn_flush_nasid_entry  {
+	struct sn_flush_device_kernel **widget_p; // Used as an array of wid_num
+	u64 iio_itte[8];
+};
+
+struct hubdev_info {
+	geoid_t				hdi_geoid;
+	short				hdi_nasid;
+	short				hdi_peer_nasid;   /* Dual Porting Peer */
+
+	struct sn_flush_nasid_entry	hdi_flush_nasid_list;
+	struct xwidget_info		hdi_xwidget_info[HUB_WIDGET_ID_MAX + 1];
+
+
+	void				*hdi_nodepda;
+	void				*hdi_node_vertex;
+	u32				max_segment_number;
+	u32				max_pcibus_number;
+};
+
+extern void hubdev_init_node(nodepda_t *, cnodeid_t);
+extern void hub_error_init(struct hubdev_info *);
+extern void ice_error_init(struct hubdev_info *);
+
+
+#endif /* _ASM_IA64_SN_XTALK_HUBDEV_H */
diff --git a/kernel/arch/ia64/sn/include/xtalk/xbow.h b/kernel/arch/ia64/sn/include/xtalk/xbow.h
new file mode 100644
index 000000000..90f37a413
--- /dev/null
+++ b/kernel/arch/ia64/sn/include/xtalk/xbow.h
@@ -0,0 +1,301 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992-1997,2000-2006 Silicon Graphics, Inc. All Rights
+ * Reserved.
+ */
+#ifndef _ASM_IA64_SN_XTALK_XBOW_H
+#define _ASM_IA64_SN_XTALK_XBOW_H
+
+#define XBOW_PORT_8	0x8
+#define XBOW_PORT_C	0xc
+#define XBOW_PORT_F	0xf
+
+#define MAX_XBOW_PORTS	8	/* number of ports on xbow chip */
+#define BASE_XBOW_PORT	XBOW_PORT_8	/* Lowest external port */
+
+#define	XBOW_CREDIT	4
+
+#define MAX_XBOW_NAME 	16
+
+/* Register set for each xbow link */
+typedef volatile struct xb_linkregs_s {
+/*
+ * we access these through synergy unswizzled space, so the address
+ * gets twiddled (i.e. references to 0x4 actually go to 0x0 and vv.)
+ * That's why we put the register first and filler second.
+ */
+	u32 link_ibf;
+	u32 filler0;	/* filler for proper alignment */
+	u32 link_control;
+	u32 filler1;
+	u32 link_status;
+	u32 filler2;
+	u32 link_arb_upper;
+	u32 filler3;
+	u32 link_arb_lower;
+	u32 filler4;
+	u32 link_status_clr;
+	u32 filler5;
+	u32 link_reset;
+	u32 filler6;
+	u32 link_aux_status;
+	u32 filler7;
+} xb_linkregs_t;
+
+typedef volatile struct xbow_s {
+	/* standard widget configuration 0x000000-0x000057 */
+	struct widget_cfg xb_widget;  /* 0x000000 */
+
+	/* helper fieldnames for accessing bridge widget */
+
+#define xb_wid_id 		xb_widget.w_id
+#define xb_wid_stat 		xb_widget.w_status
+#define xb_wid_err_upper 	xb_widget.w_err_upper_addr
+#define xb_wid_err_lower 	xb_widget.w_err_lower_addr
+#define xb_wid_control		xb_widget.w_control
+#define xb_wid_req_timeout 	xb_widget.w_req_timeout
+#define xb_wid_int_upper 	xb_widget.w_intdest_upper_addr
+#define xb_wid_int_lower 	xb_widget.w_intdest_lower_addr
+#define xb_wid_err_cmdword 	xb_widget.w_err_cmd_word
+#define xb_wid_llp 		xb_widget.w_llp_cfg
+#define xb_wid_stat_clr 	xb_widget.w_tflush
+
+/*
+ * we access these through synergy unswizzled space, so the address
+ * gets twiddled (i.e. references to 0x4 actually go to 0x0 and vv.)
+ * That's why we put the register first and filler second.
+ */
+	/* xbow-specific widget configuration    0x000058-0x0000FF */
+	u32 xb_wid_arb_reload; /* 0x00005C */
+	u32 _pad_000058;
+	u32 xb_perf_ctr_a;	/* 0x000064 */
+	u32 _pad_000060;
+	u32 xb_perf_ctr_b;	/* 0x00006c */
+	u32 _pad_000068;
+	u32 xb_nic;		/* 0x000074 */
+	u32 _pad_000070;
+
+	/* Xbridge only */
+	u32 xb_w0_rst_fnc;	/* 0x00007C */
+	u32 _pad_000078;
+	u32 xb_l8_rst_fnc;	/* 0x000084 */
+	u32 _pad_000080;
+	u32 xb_l9_rst_fnc;	/* 0x00008c */
+	u32 _pad_000088;
+	u32 xb_la_rst_fnc;	/* 0x000094 */
+	u32 _pad_000090;
+	u32 xb_lb_rst_fnc;	/* 0x00009c */
+	u32 _pad_000098;
+	u32 xb_lc_rst_fnc;	/* 0x0000a4 */
+	u32 _pad_0000a0;
+	u32 xb_ld_rst_fnc;	/* 0x0000ac */
+	u32 _pad_0000a8;
+	u32 xb_le_rst_fnc;	/* 0x0000b4 */
+	u32 _pad_0000b0;
+	u32 xb_lf_rst_fnc;	/* 0x0000bc */
+	u32 _pad_0000b8;
+	u32 xb_lock;		/* 0x0000c4 */
+	u32 _pad_0000c0;
+	u32 xb_lock_clr;	/* 0x0000cc */
+	u32 _pad_0000c8;
+	/* end of Xbridge only */
+	u32 _pad_0000d0[12];
+
+	/* Link Specific Registers, port 8..15   0x000100-0x000300 */
+	xb_linkregs_t xb_link_raw[MAX_XBOW_PORTS];
+} xbow_t;
+
+#define xb_link(p) xb_link_raw[(p) & (MAX_XBOW_PORTS - 1)]
+
+#define XB_FLAGS_EXISTS		0x1	/* device exists */
+#define XB_FLAGS_MASTER		0x2
+#define XB_FLAGS_SLAVE		0x0
+#define XB_FLAGS_GBR		0x4
+#define XB_FLAGS_16BIT		0x8
+#define XB_FLAGS_8BIT		0x0
+
+/* is widget port number valid?  (based on version 7.0 of xbow spec) */
+#define XBOW_WIDGET_IS_VALID(wid) ((wid) >= XBOW_PORT_8 && (wid) <= XBOW_PORT_F)
+
+/* whether to use upper or lower arbitration register, given source widget id */
+#define XBOW_ARB_IS_UPPER(wid) 	((wid) >= XBOW_PORT_8 && (wid) <= XBOW_PORT_B)
+#define XBOW_ARB_IS_LOWER(wid) 	((wid) >= XBOW_PORT_C && (wid) <= XBOW_PORT_F)
+
+/* offset of arbitration register, given source widget id */
+#define XBOW_ARB_OFF(wid) 	(XBOW_ARB_IS_UPPER(wid) ? 0x1c : 0x24)
+
+#define	XBOW_WID_ID		WIDGET_ID
+#define	XBOW_WID_STAT		WIDGET_STATUS
+#define	XBOW_WID_ERR_UPPER	WIDGET_ERR_UPPER_ADDR
+#define	XBOW_WID_ERR_LOWER	WIDGET_ERR_LOWER_ADDR
+#define	XBOW_WID_CONTROL	WIDGET_CONTROL
+#define	XBOW_WID_REQ_TO		WIDGET_REQ_TIMEOUT
+#define	XBOW_WID_INT_UPPER	WIDGET_INTDEST_UPPER_ADDR
+#define	XBOW_WID_INT_LOWER	WIDGET_INTDEST_LOWER_ADDR
+#define	XBOW_WID_ERR_CMDWORD	WIDGET_ERR_CMD_WORD
+#define	XBOW_WID_LLP		WIDGET_LLP_CFG
+#define	XBOW_WID_STAT_CLR	WIDGET_TFLUSH
+#define XBOW_WID_ARB_RELOAD 	0x5c
+#define XBOW_WID_PERF_CTR_A 	0x64
+#define XBOW_WID_PERF_CTR_B 	0x6c
+#define XBOW_WID_NIC 		0x74
+
+/* Xbridge only */
+#define XBOW_W0_RST_FNC		0x00007C
+#define	XBOW_L8_RST_FNC		0x000084
+#define	XBOW_L9_RST_FNC		0x00008c
+#define	XBOW_LA_RST_FNC		0x000094
+#define	XBOW_LB_RST_FNC		0x00009c
+#define	XBOW_LC_RST_FNC		0x0000a4
+#define	XBOW_LD_RST_FNC		0x0000ac
+#define	XBOW_LE_RST_FNC		0x0000b4
+#define	XBOW_LF_RST_FNC		0x0000bc
+#define XBOW_RESET_FENCE(x) ((x) > 7 && (x) < 16) ? \
+				(XBOW_W0_RST_FNC + ((x) - 7) * 8) : \
+				((x) == 0) ? XBOW_W0_RST_FNC : 0
+#define XBOW_LOCK		0x0000c4
+#define XBOW_LOCK_CLR		0x0000cc
+/* End of Xbridge only */
+
+/* used only in ide, but defined here within the reserved portion */
+/* of the widget0 address space (before 0xf4) */
+#define	XBOW_WID_UNDEF		0xe4
+
+/* xbow link register set base, legal value for x is 0x8..0xf */
+#define	XB_LINK_BASE		0x100
+#define	XB_LINK_OFFSET		0x40
+#define	XB_LINK_REG_BASE(x)	(XB_LINK_BASE + ((x) & (MAX_XBOW_PORTS - 1)) * XB_LINK_OFFSET)
+
+#define	XB_LINK_IBUF_FLUSH(x)	(XB_LINK_REG_BASE(x) + 0x4)
+#define	XB_LINK_CTRL(x)		(XB_LINK_REG_BASE(x) + 0xc)
+#define	XB_LINK_STATUS(x)	(XB_LINK_REG_BASE(x) + 0x14)
+#define	XB_LINK_ARB_UPPER(x)	(XB_LINK_REG_BASE(x) + 0x1c)
+#define	XB_LINK_ARB_LOWER(x)	(XB_LINK_REG_BASE(x) + 0x24)
+#define	XB_LINK_STATUS_CLR(x)	(XB_LINK_REG_BASE(x) + 0x2c)
+#define	XB_LINK_RESET(x)	(XB_LINK_REG_BASE(x) + 0x34)
+#define	XB_LINK_AUX_STATUS(x)	(XB_LINK_REG_BASE(x) + 0x3c)
+
+/* link_control(x) */
+#define	XB_CTRL_LINKALIVE_IE		0x80000000	/* link comes alive */
+/* reserved:			0x40000000 */
+#define	XB_CTRL_PERF_CTR_MODE_MSK	0x30000000	/* perf counter mode */
+#define	XB_CTRL_IBUF_LEVEL_MSK		0x0e000000	/* input packet buffer
+							   level */
+#define	XB_CTRL_8BIT_MODE		0x01000000	/* force link into 8
+							   bit mode */
+#define XB_CTRL_BAD_LLP_PKT		0x00800000	/* force bad LLP
+							   packet */
+#define XB_CTRL_WIDGET_CR_MSK		0x007c0000	/* LLP widget credit
+							   mask */
+#define XB_CTRL_WIDGET_CR_SHFT	18			/* LLP widget credit
+							   shift */
+#define XB_CTRL_ILLEGAL_DST_IE		0x00020000	/* illegal destination
+							 */
+#define XB_CTRL_OALLOC_IBUF_IE		0x00010000	/* overallocated input
+							   buffer */
+/* reserved:			0x0000fe00 */
+#define XB_CTRL_BNDWDTH_ALLOC_IE	0x00000100	/* bandwidth alloc */
+#define XB_CTRL_RCV_CNT_OFLOW_IE	0x00000080	/* rcv retry overflow */
+#define XB_CTRL_XMT_CNT_OFLOW_IE	0x00000040	/* xmt retry overflow */
+#define XB_CTRL_XMT_MAX_RTRY_IE		0x00000020	/* max transmit retry */
+#define XB_CTRL_RCV_IE			0x00000010	/* receive */
+#define XB_CTRL_XMT_RTRY_IE		0x00000008	/* transmit retry */
+/* reserved:			0x00000004 */
+#define	XB_CTRL_MAXREQ_TOUT_IE		0x00000002	/* maximum request
+							   timeout */
+#define	XB_CTRL_SRC_TOUT_IE		0x00000001	/* source timeout */
+
+/* link_status(x) */
+#define	XB_STAT_LINKALIVE		XB_CTRL_LINKALIVE_IE
+/* reserved:			0x7ff80000 */
+#define	XB_STAT_MULTI_ERR		0x00040000	/* multi error */
+#define	XB_STAT_ILLEGAL_DST_ERR		XB_CTRL_ILLEGAL_DST_IE
+#define	XB_STAT_OALLOC_IBUF_ERR		XB_CTRL_OALLOC_IBUF_IE
+#define	XB_STAT_BNDWDTH_ALLOC_ID_MSK	0x0000ff00	/* port bitmask */
+#define	XB_STAT_RCV_CNT_OFLOW_ERR	XB_CTRL_RCV_CNT_OFLOW_IE
+#define	XB_STAT_XMT_CNT_OFLOW_ERR	XB_CTRL_XMT_CNT_OFLOW_IE
+#define	XB_STAT_XMT_MAX_RTRY_ERR	XB_CTRL_XMT_MAX_RTRY_IE
+#define	XB_STAT_RCV_ERR			XB_CTRL_RCV_IE
+#define	XB_STAT_XMT_RTRY_ERR		XB_CTRL_XMT_RTRY_IE
+/* reserved:			0x00000004 */
+#define	XB_STAT_MAXREQ_TOUT_ERR		XB_CTRL_MAXREQ_TOUT_IE
+#define	XB_STAT_SRC_TOUT_ERR		XB_CTRL_SRC_TOUT_IE
+
+/* link_aux_status(x) */
+#define	XB_AUX_STAT_RCV_CNT	0xff000000
+#define	XB_AUX_STAT_XMT_CNT	0x00ff0000
+#define	XB_AUX_STAT_TOUT_DST	0x0000ff00
+#define	XB_AUX_LINKFAIL_RST_BAD	0x00000040
+#define	XB_AUX_STAT_PRESENT	0x00000020
+#define	XB_AUX_STAT_PORT_WIDTH	0x00000010
+/*	reserved:		0x0000000f */
+
+/*
+ * link_arb_upper/link_arb_lower(x), (reg) should be the link_arb_upper
+ * register if (x) is 0x8..0xb, link_arb_lower if (x) is 0xc..0xf
+ */
+#define	XB_ARB_GBR_MSK		0x1f
+#define	XB_ARB_RR_MSK		0x7
+#define	XB_ARB_GBR_SHFT(x)	(((x) & 0x3) * 8)
+#define	XB_ARB_RR_SHFT(x)	(((x) & 0x3) * 8 + 5)
+#define	XB_ARB_GBR_CNT(reg,x)	((reg) >> XB_ARB_GBR_SHFT(x) & XB_ARB_GBR_MSK)
+#define	XB_ARB_RR_CNT(reg,x)	((reg) >> XB_ARB_RR_SHFT(x) & XB_ARB_RR_MSK)
+
+/* XBOW_WID_STAT */
+#define	XB_WID_STAT_LINK_INTR_SHFT	(24)
+#define	XB_WID_STAT_LINK_INTR_MASK	(0xFF << XB_WID_STAT_LINK_INTR_SHFT)
+#define	XB_WID_STAT_LINK_INTR(x) \
+	(0x1 << (((x)&7) + XB_WID_STAT_LINK_INTR_SHFT))
+#define	XB_WID_STAT_WIDGET0_INTR	0x00800000
+#define XB_WID_STAT_SRCID_MASK		0x000003c0	/* Xbridge only */
+#define	XB_WID_STAT_REG_ACC_ERR		0x00000020
+#define XB_WID_STAT_RECV_TOUT		0x00000010	/* Xbridge only */
+#define XB_WID_STAT_ARB_TOUT		0x00000008	/* Xbridge only */
+#define	XB_WID_STAT_XTALK_ERR		0x00000004
+#define XB_WID_STAT_DST_TOUT		0x00000002	/* Xbridge only */
+#define	XB_WID_STAT_MULTI_ERR		0x00000001
+
+#define XB_WID_STAT_SRCID_SHFT		6
+
+/* XBOW_WID_CONTROL */
+#define XB_WID_CTRL_REG_ACC_IE		XB_WID_STAT_REG_ACC_ERR
+#define XB_WID_CTRL_RECV_TOUT		XB_WID_STAT_RECV_TOUT
+#define XB_WID_CTRL_ARB_TOUT		XB_WID_STAT_ARB_TOUT
+#define XB_WID_CTRL_XTALK_IE		XB_WID_STAT_XTALK_ERR
+
+/* XBOW_WID_INT_UPPER */
+/* defined in xwidget.h for WIDGET_INTDEST_UPPER_ADDR */
+
+/* XBOW WIDGET part number, in the ID register */
+#define XBOW_WIDGET_PART_NUM	0x0		/* crossbow */
+#define XXBOW_WIDGET_PART_NUM	0xd000		/* Xbridge */
+#define	XBOW_WIDGET_MFGR_NUM	0x0
+#define	XXBOW_WIDGET_MFGR_NUM	0x0
+#define PXBOW_WIDGET_PART_NUM   0xd100		/* PIC */
+
+#define	XBOW_REV_1_0		0x1	/* xbow rev 1.0 is "1" */
+#define	XBOW_REV_1_1		0x2	/* xbow rev 1.1 is "2" */
+#define XBOW_REV_1_2		0x3	/* xbow rev 1.2 is "3" */
+#define XBOW_REV_1_3		0x4	/* xbow rev 1.3 is "4" */
+#define XBOW_REV_2_0		0x5	/* xbow rev 2.0 is "5" */
+
+#define XXBOW_PART_REV_1_0		(XXBOW_WIDGET_PART_NUM << 4 | 0x1 )
+#define XXBOW_PART_REV_2_0		(XXBOW_WIDGET_PART_NUM << 4 | 0x2 )
+
+/* XBOW_WID_ARB_RELOAD */
+#define	XBOW_WID_ARB_RELOAD_INT	0x3f	/* GBR reload interval */
+
+#define IS_XBRIDGE_XBOW(wid) \
+	(XWIDGET_PART_NUM(wid) == XXBOW_WIDGET_PART_NUM && \
+	XWIDGET_MFG_NUM(wid) == XXBOW_WIDGET_MFGR_NUM)
+
+#define IS_PIC_XBOW(wid) \
+	(XWIDGET_PART_NUM(wid) == PXBOW_WIDGET_PART_NUM && \
+	XWIDGET_MFG_NUM(wid) == XXBOW_WIDGET_MFGR_NUM)
+
+#define XBOW_WAR_ENABLED(pv, widid) ((1 << XWIDGET_REV_NUM(widid)) & pv)
+
+#endif /* _ASM_IA64_SN_XTALK_XBOW_H */
diff --git a/kernel/arch/ia64/sn/include/xtalk/xwidgetdev.h b/kernel/arch/ia64/sn/include/xtalk/xwidgetdev.h
new file mode 100644
index 000000000..2800eda0f
--- /dev/null
+++ b/kernel/arch/ia64/sn/include/xtalk/xwidgetdev.h
@@ -0,0 +1,70 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992-1997,2000-2003 Silicon Graphics, Inc. All Rights Reserved.
+ */
+#ifndef _ASM_IA64_SN_XTALK_XWIDGET_H
+#define _ASM_IA64_SN_XTALK_XWIDGET_H
+
+/* WIDGET_ID */
+#define WIDGET_REV_NUM                  0xf0000000
+#define WIDGET_PART_NUM                 0x0ffff000
+#define WIDGET_MFG_NUM                  0x00000ffe
+#define WIDGET_REV_NUM_SHFT             28
+#define WIDGET_PART_NUM_SHFT            12
+#define WIDGET_MFG_NUM_SHFT             1
+
+#define XWIDGET_PART_NUM(widgetid) (((widgetid) & WIDGET_PART_NUM) >> WIDGET_PART_NUM_SHFT)
+#define XWIDGET_REV_NUM(widgetid) (((widgetid) & WIDGET_REV_NUM) >> WIDGET_REV_NUM_SHFT)
+#define XWIDGET_MFG_NUM(widgetid) (((widgetid) & WIDGET_MFG_NUM) >> WIDGET_MFG_NUM_SHFT)
+#define XWIDGET_PART_REV_NUM(widgetid) ((XWIDGET_PART_NUM(widgetid) << 4) | \
+                                        XWIDGET_REV_NUM(widgetid))
+#define XWIDGET_PART_REV_NUM_REV(partrev) (partrev & 0xf)
+
+/* widget configuration registers */
+struct widget_cfg{
+	u32	w_id;	/* 0x04 */
+	u32	w_pad_0;	/* 0x00 */
+	u32	w_status;	/* 0x0c */
+	u32	w_pad_1;	/* 0x08 */
+	u32	w_err_upper_addr;	/* 0x14 */
+	u32	w_pad_2;	/* 0x10 */
+	u32	w_err_lower_addr;	/* 0x1c */
+	u32	w_pad_3;	/* 0x18 */
+	u32	w_control;	/* 0x24 */
+	u32	w_pad_4;	/* 0x20 */
+	u32	w_req_timeout;	/* 0x2c */
+	u32	w_pad_5;	/* 0x28 */
+	u32	w_intdest_upper_addr;	/* 0x34 */
+	u32	w_pad_6;	/* 0x30 */
+	u32	w_intdest_lower_addr;	/* 0x3c */
+	u32	w_pad_7;	/* 0x38 */
+	u32	w_err_cmd_word;	/* 0x44 */
+	u32	w_pad_8;	/* 0x40 */
+	u32	w_llp_cfg;	/* 0x4c */
+	u32	w_pad_9;	/* 0x48 */
+	u32	w_tflush;	/* 0x54 */
+	u32	w_pad_10;	/* 0x50 */
+};
+
+/*
+ * Crosstalk Widget Hardware Identification, as defined in the Crosstalk spec.
+ */
+struct xwidget_hwid{
+	int		mfg_num;
+	int		rev_num;
+	int		part_num;
+};
+
+struct xwidget_info{
+
+	struct xwidget_hwid	xwi_hwid;	/* Widget Identification */
+	char			xwi_masterxid;	/* Hub's Widget Port Number */
+	void			*xwi_hubinfo;     /* Hub's provider private info */
+	u64			*xwi_hub_provider; /* prom provider functions */
+	void			*xwi_vertex;
+};
+
+#endif                          /* _ASM_IA64_SN_XTALK_XWIDGET_H */
diff --git a/kernel/arch/ia64/sn/kernel/Makefile b/kernel/arch/ia64/sn/kernel/Makefile
new file mode 100644
index 000000000..d27df1d45
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/Makefile
@@ -0,0 +1,18 @@
+# arch/ia64/sn/kernel/Makefile
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# Copyright (C) 1999,2001-2006,2008 Silicon Graphics, Inc.  All Rights Reserved.
+#
+
+ccflags-y := -Iarch/ia64/sn/include
+
+obj-y				+= setup.o bte.o bte_error.o irq.o mca.o idle.o \
+				   huberror.o io_acpi_init.o io_common.o \
+				   io_init.o iomv.o klconflib.o pio_phys.o \
+				   sn2/
+obj-$(CONFIG_IA64_GENERIC)      += machvec.o
+obj-$(CONFIG_SGI_TIOCX)		+= tiocx.o
+obj-$(CONFIG_PCI_MSI)		+= msi_sn.o
diff --git a/kernel/arch/ia64/sn/kernel/bte.c b/kernel/arch/ia64/sn/kernel/bte.c
new file mode 100644
index 000000000..b2eb48490
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/bte.c
@@ -0,0 +1,471 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2000-2007 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/module.h>
+#include <asm/sn/nodepda.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/arch.h>
+#include <asm/sn/sn_cpuid.h>
+#include <asm/sn/pda.h>
+#include <asm/sn/shubio.h>
+#include <asm/nodedata.h>
+#include <asm/delay.h>
+
+#include <linux/bootmem.h>
+#include <linux/string.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+
+#include <asm/sn/bte.h>
+
+#ifndef L1_CACHE_MASK
+#define L1_CACHE_MASK (L1_CACHE_BYTES - 1)
+#endif
+
+/* two interfaces on two btes */
+#define MAX_INTERFACES_TO_TRY		4
+#define MAX_NODES_TO_TRY		2
+
+static struct bteinfo_s *bte_if_on_node(nasid_t nasid, int interface)
+{
+	nodepda_t *tmp_nodepda;
+
+	if (nasid_to_cnodeid(nasid) == -1)
+		return (struct bteinfo_s *)NULL;
+
+	tmp_nodepda = NODEPDA(nasid_to_cnodeid(nasid));
+	return &tmp_nodepda->bte_if[interface];
+
+}
+
+static inline void bte_start_transfer(struct bteinfo_s *bte, u64 len, u64 mode)
+{
+	if (is_shub2()) {
+		BTE_CTRL_STORE(bte, (IBLS_BUSY | ((len) | (mode) << 24)));
+	} else {
+		BTE_LNSTAT_STORE(bte, len);
+		BTE_CTRL_STORE(bte, mode);
+	}
+}
+
+/************************************************************************
+ * Block Transfer Engine copy related functions.
+ *
+ ***********************************************************************/
+
+/*
+ * bte_copy(src, dest, len, mode, notification)
+ *
+ * Use the block transfer engine to move kernel memory from src to dest
+ * using the assigned mode.
+ *
+ * Parameters:
+ *   src - physical address of the transfer source.
+ *   dest - physical address of the transfer destination.
+ *   len - number of bytes to transfer from source to dest.
+ *   mode - hardware defined.  See reference information
+ *          for IBCT0/1 in the SHUB Programmers Reference
+ *   notification - kernel virtual address of the notification cache
+ *                  line.  If NULL, the default is used and
+ *                  the bte_copy is synchronous.
+ *
+ * NOTE:  This function requires src, dest, and len to
+ * be cacheline aligned.
+ */
+bte_result_t bte_copy(u64 src, u64 dest, u64 len, u64 mode, void *notification)
+{
+	u64 transfer_size;
+	u64 transfer_stat;
+	u64 notif_phys_addr;
+	struct bteinfo_s *bte;
+	bte_result_t bte_status;
+	unsigned long irq_flags;
+	unsigned long itc_end = 0;
+	int nasid_to_try[MAX_NODES_TO_TRY];
+	int my_nasid = cpuid_to_nasid(raw_smp_processor_id());
+	int bte_if_index, nasid_index;
+	int bte_first, btes_per_node = BTES_PER_NODE;
+
+	BTE_PRINTK(("bte_copy(0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%p)\n",
+		    src, dest, len, mode, notification));
+
+	if (len == 0) {
+		return BTE_SUCCESS;
+	}
+
+	BUG_ON(len & L1_CACHE_MASK);
+	BUG_ON(src & L1_CACHE_MASK);
+	BUG_ON(dest & L1_CACHE_MASK);
+	BUG_ON(len > BTE_MAX_XFER);
+
+	/*
+	 * Start with interface corresponding to cpu number
+	 */
+	bte_first = raw_smp_processor_id() % btes_per_node;
+
+	if (mode & BTE_USE_DEST) {
+		/* try remote then local */
+		nasid_to_try[0] = NASID_GET(dest);
+		if (mode & BTE_USE_ANY) {
+			nasid_to_try[1] = my_nasid;
+		} else {
+			nasid_to_try[1] = 0;
+		}
+	} else {
+		/* try local then remote */
+		nasid_to_try[0] = my_nasid;
+		if (mode & BTE_USE_ANY) {
+			nasid_to_try[1] = NASID_GET(dest);
+		} else {
+			nasid_to_try[1] = 0;
+		}
+	}
+
+retry_bteop:
+	do {
+		local_irq_save(irq_flags);
+
+		bte_if_index = bte_first;
+		nasid_index = 0;
+
+		/* Attempt to lock one of the BTE interfaces. */
+		while (nasid_index < MAX_NODES_TO_TRY) {
+			bte = bte_if_on_node(nasid_to_try[nasid_index],bte_if_index);
+
+			if (bte == NULL) {
+				nasid_index++;
+				continue;
+			}
+
+			if (spin_trylock(&bte->spinlock)) {
+				if (!(*bte->most_rcnt_na & BTE_WORD_AVAILABLE) ||
+				    (BTE_LNSTAT_LOAD(bte) & BTE_ACTIVE)) {
+					/* Got the lock but BTE still busy */
+					spin_unlock(&bte->spinlock);
+				} else {
+					/* we got the lock and it's not busy */
+					break;
+				}
+			}
+
+			bte_if_index = (bte_if_index + 1) % btes_per_node; /* Next interface */
+			if (bte_if_index == bte_first) {
+				/*
+				 * We've tried all interfaces on this node
+				 */
+				nasid_index++;
+			}
+
+			bte = NULL;
+		}
+
+		if (bte != NULL) {
+			break;
+		}
+
+		local_irq_restore(irq_flags);
+
+		if (!(mode & BTE_WACQUIRE)) {
+			return BTEFAIL_NOTAVAIL;
+		}
+	} while (1);
+
+	if (notification == NULL) {
+		/* User does not want to be notified. */
+		bte->most_rcnt_na = &bte->notify;
+	} else {
+		bte->most_rcnt_na = notification;
+	}
+
+	/* Calculate the number of cache lines to transfer. */
+	transfer_size = ((len >> L1_CACHE_SHIFT) & BTE_LEN_MASK);
+
+	/* Initialize the notification to a known value. */
+	*bte->most_rcnt_na = BTE_WORD_BUSY;
+	notif_phys_addr = (u64)bte->most_rcnt_na;
+
+	/* Set the source and destination registers */
+	BTE_PRINTKV(("IBSA = 0x%lx)\n", src));
+	BTE_SRC_STORE(bte, src);
+	BTE_PRINTKV(("IBDA = 0x%lx)\n", dest));
+	BTE_DEST_STORE(bte, dest);
+
+	/* Set the notification register */
+	BTE_PRINTKV(("IBNA = 0x%lx)\n", notif_phys_addr));
+	BTE_NOTIF_STORE(bte, notif_phys_addr);
+
+	/* Initiate the transfer */
+	BTE_PRINTK(("IBCT = 0x%lx)\n", BTE_VALID_MODE(mode)));
+	bte_start_transfer(bte, transfer_size, BTE_VALID_MODE(mode));
+
+	itc_end = ia64_get_itc() + (40000000 * local_cpu_data->cyc_per_usec);
+
+	spin_unlock_irqrestore(&bte->spinlock, irq_flags);
+
+	if (notification != NULL) {
+		return BTE_SUCCESS;
+	}
+
+	while ((transfer_stat = *bte->most_rcnt_na) == BTE_WORD_BUSY) {
+		cpu_relax();
+		if (ia64_get_itc() > itc_end) {
+			BTE_PRINTK(("BTE timeout nasid 0x%x bte%d IBLS = 0x%lx na 0x%lx\n",
+				NASID_GET(bte->bte_base_addr), bte->bte_num,
+				BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na) );
+			bte->bte_error_count++;
+			bte->bh_error = IBLS_ERROR;
+			bte_error_handler((unsigned long)NODEPDA(bte->bte_cnode));
+			*bte->most_rcnt_na = BTE_WORD_AVAILABLE;
+			goto retry_bteop;
+		}
+	}
+
+	BTE_PRINTKV((" Delay Done.  IBLS = 0x%lx, most_rcnt_na = 0x%lx\n",
+		     BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
+
+	if (transfer_stat & IBLS_ERROR) {
+		bte_status = BTE_GET_ERROR_STATUS(transfer_stat);
+	} else {
+		bte_status = BTE_SUCCESS;
+	}
+	*bte->most_rcnt_na = BTE_WORD_AVAILABLE;
+
+	BTE_PRINTK(("Returning status is 0x%lx and most_rcnt_na is 0x%lx\n",
+		    BTE_LNSTAT_LOAD(bte), *bte->most_rcnt_na));
+
+	return bte_status;
+}
+
+EXPORT_SYMBOL(bte_copy);
+
+/*
+ * bte_unaligned_copy(src, dest, len, mode)
+ *
+ * use the block transfer engine to move kernel
+ * memory from src to dest using the assigned mode.
+ *
+ * Parameters:
+ *   src - physical address of the transfer source.
+ *   dest - physical address of the transfer destination.
+ *   len - number of bytes to transfer from source to dest.
+ *   mode - hardware defined.  See reference information
+ *          for IBCT0/1 in the SGI documentation.
+ *
+ * NOTE: If the source, dest, and len are all cache line aligned,
+ * then it would be _FAR_ preferable to use bte_copy instead.
+ */
+bte_result_t bte_unaligned_copy(u64 src, u64 dest, u64 len, u64 mode)
+{
+	int destFirstCacheOffset;
+	u64 headBteSource;
+	u64 headBteLen;
+	u64 headBcopySrcOffset;
+	u64 headBcopyDest;
+	u64 headBcopyLen;
+	u64 footBteSource;
+	u64 footBteLen;
+	u64 footBcopyDest;
+	u64 footBcopyLen;
+	bte_result_t rv;
+	char *bteBlock, *bteBlock_unaligned;
+
+	if (len == 0) {
+		return BTE_SUCCESS;
+	}
+
+	/* temporary buffer used during unaligned transfers */
+	bteBlock_unaligned = kmalloc(len + 3 * L1_CACHE_BYTES, GFP_KERNEL);
+	if (bteBlock_unaligned == NULL) {
+		return BTEFAIL_NOTAVAIL;
+	}
+	bteBlock = (char *)L1_CACHE_ALIGN((u64) bteBlock_unaligned);
+
+	headBcopySrcOffset = src & L1_CACHE_MASK;
+	destFirstCacheOffset = dest & L1_CACHE_MASK;
+
+	/*
+	 * At this point, the transfer is broken into
+	 * (up to) three sections.  The first section is
+	 * from the start address to the first physical
+	 * cache line, the second is from the first physical
+	 * cache line to the last complete cache line,
+	 * and the third is from the last cache line to the
+	 * end of the buffer.  The first and third sections
+	 * are handled by bte copying into a temporary buffer
+	 * and then bcopy'ing the necessary section into the
+	 * final location.  The middle section is handled with
+	 * a standard bte copy.
+	 *
+	 * One nasty exception to the above rule is when the
+	 * source and destination are not symmetrically
+	 * mis-aligned.  If the source offset from the first
+	 * cache line is different from the destination offset,
+	 * we make the first section be the entire transfer
+	 * and the bcopy the entire block into place.
+	 */
+	if (headBcopySrcOffset == destFirstCacheOffset) {
+
+		/*
+		 * Both the source and destination are the same
+		 * distance from a cache line boundary so we can
+		 * use the bte to transfer the bulk of the
+		 * data.
+		 */
+		headBteSource = src & ~L1_CACHE_MASK;
+		headBcopyDest = dest;
+		if (headBcopySrcOffset) {
+			headBcopyLen =
+			    (len >
+			     (L1_CACHE_BYTES -
+			      headBcopySrcOffset) ? L1_CACHE_BYTES
+			     - headBcopySrcOffset : len);
+			headBteLen = L1_CACHE_BYTES;
+		} else {
+			headBcopyLen = 0;
+			headBteLen = 0;
+		}
+
+		if (len > headBcopyLen) {
+			footBcopyLen = (len - headBcopyLen) & L1_CACHE_MASK;
+			footBteLen = L1_CACHE_BYTES;
+
+			footBteSource = src + len - footBcopyLen;
+			footBcopyDest = dest + len - footBcopyLen;
+
+			if (footBcopyDest == (headBcopyDest + headBcopyLen)) {
+				/*
+				 * We have two contiguous bcopy
+				 * blocks.  Merge them.
+				 */
+				headBcopyLen += footBcopyLen;
+				headBteLen += footBteLen;
+			} else if (footBcopyLen > 0) {
+				rv = bte_copy(footBteSource,
+					      ia64_tpa((unsigned long)bteBlock),
+					      footBteLen, mode, NULL);
+				if (rv != BTE_SUCCESS) {
+					kfree(bteBlock_unaligned);
+					return rv;
+				}
+
+				memcpy(__va(footBcopyDest),
+				       (char *)bteBlock, footBcopyLen);
+			}
+		} else {
+			footBcopyLen = 0;
+			footBteLen = 0;
+		}
+
+		if (len > (headBcopyLen + footBcopyLen)) {
+			/* now transfer the middle. */
+			rv = bte_copy((src + headBcopyLen),
+				      (dest +
+				       headBcopyLen),
+				      (len - headBcopyLen -
+				       footBcopyLen), mode, NULL);
+			if (rv != BTE_SUCCESS) {
+				kfree(bteBlock_unaligned);
+				return rv;
+			}
+
+		}
+	} else {
+
+		/*
+		 * The transfer is not symmetric, we will
+		 * allocate a buffer large enough for all the
+		 * data, bte_copy into that buffer and then
+		 * bcopy to the destination.
+		 */
+
+		headBcopySrcOffset = src & L1_CACHE_MASK;
+		headBcopyDest = dest;
+		headBcopyLen = len;
+
+		headBteSource = src - headBcopySrcOffset;
+		/* Add the leading and trailing bytes from source */
+		headBteLen = L1_CACHE_ALIGN(len + headBcopySrcOffset);
+	}
+
+	if (headBcopyLen > 0) {
+		rv = bte_copy(headBteSource,
+			      ia64_tpa((unsigned long)bteBlock), headBteLen,
+			      mode, NULL);
+		if (rv != BTE_SUCCESS) {
+			kfree(bteBlock_unaligned);
+			return rv;
+		}
+
+		memcpy(__va(headBcopyDest), ((char *)bteBlock +
+					     headBcopySrcOffset), headBcopyLen);
+	}
+	kfree(bteBlock_unaligned);
+	return BTE_SUCCESS;
+}
+
+EXPORT_SYMBOL(bte_unaligned_copy);
+
+/************************************************************************
+ * Block Transfer Engine initialization functions.
+ *
+ ***********************************************************************/
+
+/*
+ * bte_init_node(nodepda, cnode)
+ *
+ * Initialize the nodepda structure with BTE base addresses and
+ * spinlocks.
+ */
+void bte_init_node(nodepda_t * mynodepda, cnodeid_t cnode)
+{
+	int i;
+
+	/*
+	 * Indicate that all the block transfer engines on this node
+	 * are available.
+	 */
+
+	/*
+	 * Allocate one bte_recover_t structure per node.  It holds
+	 * the recovery lock for node.  All the bte interface structures
+	 * will point at this one bte_recover structure to get the lock.
+	 */
+	spin_lock_init(&mynodepda->bte_recovery_lock);
+	init_timer(&mynodepda->bte_recovery_timer);
+	mynodepda->bte_recovery_timer.function = bte_error_handler;
+	mynodepda->bte_recovery_timer.data = (unsigned long)mynodepda;
+
+	for (i = 0; i < BTES_PER_NODE; i++) {
+		u64 *base_addr;
+
+		/* Which link status register should we use? */
+		base_addr = (u64 *)
+		    REMOTE_HUB_ADDR(cnodeid_to_nasid(cnode), BTE_BASE_ADDR(i));
+		mynodepda->bte_if[i].bte_base_addr = base_addr;
+		mynodepda->bte_if[i].bte_source_addr = BTE_SOURCE_ADDR(base_addr);
+		mynodepda->bte_if[i].bte_destination_addr = BTE_DEST_ADDR(base_addr);
+		mynodepda->bte_if[i].bte_control_addr = BTE_CTRL_ADDR(base_addr);
+		mynodepda->bte_if[i].bte_notify_addr = BTE_NOTIF_ADDR(base_addr);
+
+		/*
+		 * Initialize the notification and spinlock
+		 * so the first transfer can occur.
+		 */
+		mynodepda->bte_if[i].most_rcnt_na =
+		    &(mynodepda->bte_if[i].notify);
+		mynodepda->bte_if[i].notify = BTE_WORD_AVAILABLE;
+		spin_lock_init(&mynodepda->bte_if[i].spinlock);
+
+		mynodepda->bte_if[i].bte_cnode = cnode;
+		mynodepda->bte_if[i].bte_error_count = 0;
+		mynodepda->bte_if[i].bte_num = i;
+		mynodepda->bte_if[i].cleanup_active = 0;
+		mynodepda->bte_if[i].bh_error = 0;
+	}
+
+}
diff --git a/kernel/arch/ia64/sn/kernel/bte_error.c b/kernel/arch/ia64/sn/kernel/bte_error.c
new file mode 100644
index 000000000..4cb09f3f1
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/bte_error.c
@@ -0,0 +1,260 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2000-2007 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/types.h>
+#include <asm/sn/sn_sal.h>
+#include "ioerror.h"
+#include <asm/sn/addrs.h>
+#include <asm/sn/shubio.h>
+#include <asm/sn/geo.h>
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+#include <asm/sn/bte.h>
+#include <asm/param.h>
+
+/*
+ * Bte error handling is done in two parts.  The first captures
+ * any crb related errors.  Since there can be multiple crbs per
+ * interface and multiple interfaces active, we need to wait until
+ * all active crbs are completed.  This is the first job of the
+ * second part error handler.  When all bte related CRBs are cleanly
+ * completed, it resets the interfaces and gets them ready for new
+ * transfers to be queued.
+ */
+
+void bte_error_handler(unsigned long);
+
+/*
+ * Wait until all BTE related CRBs are completed
+ * and then reset the interfaces.
+ */
+int shub1_bte_error_handler(unsigned long _nodepda)
+{
+	struct nodepda_s *err_nodepda = (struct nodepda_s *)_nodepda;
+	struct timer_list *recovery_timer = &err_nodepda->bte_recovery_timer;
+	nasid_t nasid;
+	int i;
+	int valid_crbs;
+	ii_imem_u_t imem;	/* II IMEM Register */
+	ii_icrb0_d_u_t icrbd;	/* II CRB Register D */
+	ii_ibcr_u_t ibcr;
+	ii_icmr_u_t icmr;
+	ii_ieclr_u_t ieclr;
+
+	BTE_PRINTK(("shub1_bte_error_handler(%p) - %d\n", err_nodepda,
+		    smp_processor_id()));
+
+	if ((err_nodepda->bte_if[0].bh_error == BTE_SUCCESS) &&
+	    (err_nodepda->bte_if[1].bh_error == BTE_SUCCESS)) {
+		BTE_PRINTK(("eh:%p:%d Nothing to do.\n", err_nodepda,
+			    smp_processor_id()));
+		return 1;
+	}
+
+	/* Determine information about our hub */
+	nasid = cnodeid_to_nasid(err_nodepda->bte_if[0].bte_cnode);
+
+	/*
+	 * A BTE transfer can use multiple CRBs.  We need to make sure
+	 * that all the BTE CRBs are complete (or timed out) before
+	 * attempting to clean up the error.  Resetting the BTE while
+	 * there are still BTE CRBs active will hang the BTE.
+	 * We should look at all the CRBs to see if they are allocated
+	 * to the BTE and see if they are still active.  When none
+	 * are active, we can continue with the cleanup.
+	 *
+	 * We also want to make sure that the local NI port is up.
+	 * When a router resets the NI port can go down, while it
+	 * goes through the LLP handshake, but then comes back up.
+	 */
+	icmr.ii_icmr_regval = REMOTE_HUB_L(nasid, IIO_ICMR);
+	if (icmr.ii_icmr_fld_s.i_crb_mark != 0) {
+		/*
+		 * There are errors which still need to be cleaned up by
+		 * hubiio_crb_error_handler
+		 */
+		mod_timer(recovery_timer, jiffies + (HZ * 5));
+		BTE_PRINTK(("eh:%p:%d Marked Giving up\n", err_nodepda,
+			    smp_processor_id()));
+		return 1;
+	}
+	if (icmr.ii_icmr_fld_s.i_crb_vld != 0) {
+
+		valid_crbs = icmr.ii_icmr_fld_s.i_crb_vld;
+
+		for (i = 0; i < IIO_NUM_CRBS; i++) {
+			if (!((1 << i) & valid_crbs)) {
+				/* This crb was not marked as valid, ignore */
+				continue;
+			}
+			icrbd.ii_icrb0_d_regval =
+			    REMOTE_HUB_L(nasid, IIO_ICRB_D(i));
+			if (icrbd.d_bteop) {
+				mod_timer(recovery_timer, jiffies + (HZ * 5));
+				BTE_PRINTK(("eh:%p:%d Valid %d, Giving up\n",
+					    err_nodepda, smp_processor_id(),
+					    i));
+				return 1;
+			}
+		}
+	}
+
+	BTE_PRINTK(("eh:%p:%d Cleaning up\n", err_nodepda, smp_processor_id()));
+	/* Re-enable both bte interfaces */
+	imem.ii_imem_regval = REMOTE_HUB_L(nasid, IIO_IMEM);
+	imem.ii_imem_fld_s.i_b0_esd = imem.ii_imem_fld_s.i_b1_esd = 1;
+	REMOTE_HUB_S(nasid, IIO_IMEM, imem.ii_imem_regval);
+
+	/* Clear BTE0/1 error bits */
+	ieclr.ii_ieclr_regval = 0;
+	if (err_nodepda->bte_if[0].bh_error != BTE_SUCCESS)
+		ieclr.ii_ieclr_fld_s.i_e_bte_0 = 1;
+	if (err_nodepda->bte_if[1].bh_error != BTE_SUCCESS)
+		ieclr.ii_ieclr_fld_s.i_e_bte_1 = 1;
+	REMOTE_HUB_S(nasid, IIO_IECLR, ieclr.ii_ieclr_regval);
+
+	/* Reinitialize both BTE state machines. */
+	ibcr.ii_ibcr_regval = REMOTE_HUB_L(nasid, IIO_IBCR);
+	ibcr.ii_ibcr_fld_s.i_soft_reset = 1;
+	REMOTE_HUB_S(nasid, IIO_IBCR, ibcr.ii_ibcr_regval);
+
+	del_timer(recovery_timer);
+	return 0;
+}
+
+/*
+ * Wait until all BTE related CRBs are completed
+ * and then reset the interfaces.
+ */
+int shub2_bte_error_handler(unsigned long _nodepda)
+{
+	struct nodepda_s *err_nodepda = (struct nodepda_s *)_nodepda;
+	struct timer_list *recovery_timer = &err_nodepda->bte_recovery_timer;
+	struct bteinfo_s *bte;
+	nasid_t nasid;
+	u64 status;
+	int i;
+
+	nasid = cnodeid_to_nasid(err_nodepda->bte_if[0].bte_cnode);
+
+	/*
+	 * Verify that all the BTEs are complete
+	 */
+	for (i = 0; i < BTES_PER_NODE; i++) {
+		bte = &err_nodepda->bte_if[i];
+		status = BTE_LNSTAT_LOAD(bte);
+		if (status & IBLS_ERROR) {
+			bte->bh_error = BTE_SHUB2_ERROR(status);
+			continue;
+		}
+		if (!(status & IBLS_BUSY))
+			continue;
+		mod_timer(recovery_timer, jiffies + (HZ * 5));
+		BTE_PRINTK(("eh:%p:%d Marked Giving up\n", err_nodepda,
+			    smp_processor_id()));
+		return 1;
+	}
+	if (ia64_sn_bte_recovery(nasid))
+		panic("bte_error_handler(): Fatal BTE Error");
+
+	del_timer(recovery_timer);
+	return 0;
+}
+
+/*
+ * Wait until all BTE related CRBs are completed
+ * and then reset the interfaces.
+ */
+void bte_error_handler(unsigned long _nodepda)
+{
+	struct nodepda_s *err_nodepda = (struct nodepda_s *)_nodepda;
+	spinlock_t *recovery_lock = &err_nodepda->bte_recovery_lock;
+	int i;
+	unsigned long irq_flags;
+	volatile u64 *notify;
+	bte_result_t bh_error;
+
+	BTE_PRINTK(("bte_error_handler(%p) - %d\n", err_nodepda,
+		    smp_processor_id()));
+
+	spin_lock_irqsave(recovery_lock, irq_flags);
+
+	/*
+	 * Lock all interfaces on this node to prevent new transfers
+	 * from being queued.
+	 */
+	for (i = 0; i < BTES_PER_NODE; i++) {
+		if (err_nodepda->bte_if[i].cleanup_active) {
+			continue;
+		}
+		spin_lock(&err_nodepda->bte_if[i].spinlock);
+		BTE_PRINTK(("eh:%p:%d locked %d\n", err_nodepda,
+			    smp_processor_id(), i));
+		err_nodepda->bte_if[i].cleanup_active = 1;
+	}
+
+	if (is_shub1()) {
+		if (shub1_bte_error_handler(_nodepda)) {
+			spin_unlock_irqrestore(recovery_lock, irq_flags);
+			return;
+		}
+	} else {
+		if (shub2_bte_error_handler(_nodepda)) {
+			spin_unlock_irqrestore(recovery_lock, irq_flags);
+			return;
+		}
+	}
+
+	for (i = 0; i < BTES_PER_NODE; i++) {
+		bh_error = err_nodepda->bte_if[i].bh_error;
+		if (bh_error != BTE_SUCCESS) {
+			/* There is an error which needs to be notified */
+			notify = err_nodepda->bte_if[i].most_rcnt_na;
+			BTE_PRINTK(("cnode %d bte %d error=0x%lx\n",
+				    err_nodepda->bte_if[i].bte_cnode,
+				    err_nodepda->bte_if[i].bte_num,
+				    IBLS_ERROR | (u64) bh_error));
+			*notify = IBLS_ERROR | bh_error;
+			err_nodepda->bte_if[i].bh_error = BTE_SUCCESS;
+		}
+
+		err_nodepda->bte_if[i].cleanup_active = 0;
+		BTE_PRINTK(("eh:%p:%d Unlocked %d\n", err_nodepda,
+			    smp_processor_id(), i));
+		spin_unlock(&err_nodepda->bte_if[i].spinlock);
+	}
+
+	spin_unlock_irqrestore(recovery_lock, irq_flags);
+}
+
+/*
+ * First part error handler.  This is called whenever any error CRB interrupt
+ * is generated by the II.
+ */
+void
+bte_crb_error_handler(cnodeid_t cnode, int btenum,
+                      int crbnum, ioerror_t * ioe, int bteop)
+{
+	struct bteinfo_s *bte;
+
+
+	bte = &(NODEPDA(cnode)->bte_if[btenum]);
+
+	/*
+	 * The caller has already figured out the error type, we save that
+	 * in the bte handle structure for the thread exercising the
+	 * interface to consume.
+	 */
+	bte->bh_error = ioe->ie_errortype + BTEFAIL_OFFSET;
+	bte->bte_error_count++;
+
+	BTE_PRINTK(("Got an error on cnode %d bte %d: HW error type 0x%x\n",
+		bte->bte_cnode, bte->bte_num, ioe->ie_errortype));
+	bte_error_handler((unsigned long) NODEPDA(cnode));
+}
+
diff --git a/kernel/arch/ia64/sn/kernel/huberror.c b/kernel/arch/ia64/sn/kernel/huberror.c
new file mode 100644
index 000000000..f925dec2d
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/huberror.c
@@ -0,0 +1,220 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992 - 1997, 2000,2002-2007 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/interrupt.h>
+#include <asm/delay.h>
+#include <asm/sn/sn_sal.h>
+#include "ioerror.h"
+#include <asm/sn/addrs.h>
+#include <asm/sn/shubio.h>
+#include <asm/sn/geo.h>
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+#include <asm/sn/bte.h>
+
+void hubiio_crb_error_handler(struct hubdev_info *hubdev_info);
+extern void bte_crb_error_handler(cnodeid_t, int, int, ioerror_t *,
+				  int);
+static irqreturn_t hub_eint_handler(int irq, void *arg)
+{
+	struct hubdev_info *hubdev_info;
+	struct ia64_sal_retval ret_stuff;
+	nasid_t nasid;
+
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+	hubdev_info = (struct hubdev_info *)arg;
+	nasid = hubdev_info->hdi_nasid;
+
+	if (is_shub1()) {
+		SAL_CALL_NOLOCK(ret_stuff, SN_SAL_HUB_ERROR_INTERRUPT,
+			(u64) nasid, 0, 0, 0, 0, 0, 0);
+
+		if ((int)ret_stuff.v0)
+			panic("%s: Fatal %s Error", __func__,
+				((nasid & 1) ? "TIO" : "HUBII"));
+
+		if (!(nasid & 1)) /* Not a TIO, handle CRB errors */
+			(void)hubiio_crb_error_handler(hubdev_info);
+	} else
+		if (nasid & 1) {	/* TIO errors */
+			SAL_CALL_NOLOCK(ret_stuff, SN_SAL_HUB_ERROR_INTERRUPT,
+				(u64) nasid, 0, 0, 0, 0, 0, 0);
+
+			if ((int)ret_stuff.v0)
+				panic("%s: Fatal TIO Error", __func__);
+		} else
+			bte_error_handler((unsigned long)NODEPDA(nasid_to_cnodeid(nasid)));
+
+	return IRQ_HANDLED;
+}
+
+/*
+ * Free the hub CRB "crbnum" which encountered an error.
+ * Assumption is, error handling was successfully done,
+ * and we now want to return the CRB back to Hub for normal usage.
+ *
+ * In order to free the CRB, all that's needed is to de-allocate it
+ *
+ * Assumption:
+ *      No other processor is mucking around with the hub control register.
+ *      So, upper layer has to single thread this.
+ */
+void hubiio_crb_free(struct hubdev_info *hubdev_info, int crbnum)
+{
+	ii_icrb0_b_u_t icrbb;
+
+	/*
+	 * The hardware does NOT clear the mark bit, so it must get cleared
+	 * here to be sure the error is not processed twice.
+	 */
+	icrbb.ii_icrb0_b_regval = REMOTE_HUB_L(hubdev_info->hdi_nasid,
+					       IIO_ICRB_B(crbnum));
+	icrbb.b_mark = 0;
+	REMOTE_HUB_S(hubdev_info->hdi_nasid, IIO_ICRB_B(crbnum),
+		     icrbb.ii_icrb0_b_regval);
+	/*
+	 * Deallocate the register wait till hub indicates it's done.
+	 */
+	REMOTE_HUB_S(hubdev_info->hdi_nasid, IIO_ICDR, (IIO_ICDR_PND | crbnum));
+	while (REMOTE_HUB_L(hubdev_info->hdi_nasid, IIO_ICDR) & IIO_ICDR_PND)
+		cpu_relax();
+
+}
+
+/*
+ * hubiio_crb_error_handler
+ *
+ *	This routine gets invoked when a hub gets an error 
+ *	interrupt. So, the routine is running in interrupt context
+ *	at error interrupt level.
+ * Action:
+ *	It's responsible for identifying ALL the CRBs that are marked
+ *	with error, and process them. 
+ *	
+ * 	If you find the CRB that's marked with error, map this to the
+ *	reason it caused error, and invoke appropriate error handler.
+ *
+ *	XXX Be aware of the information in the context register.
+ *
+ * NOTE:
+ *	Use REMOTE_HUB_* macro instead of LOCAL_HUB_* so that the interrupt
+ *	handler can be run on any node. (not necessarily the node 
+ *	corresponding to the hub that encountered error).
+ */
+
+void hubiio_crb_error_handler(struct hubdev_info *hubdev_info)
+{
+	nasid_t nasid;
+	ii_icrb0_a_u_t icrba;	/* II CRB Register A */
+	ii_icrb0_b_u_t icrbb;	/* II CRB Register B */
+	ii_icrb0_c_u_t icrbc;	/* II CRB Register C */
+	ii_icrb0_d_u_t icrbd;	/* II CRB Register D */
+	ii_icrb0_e_u_t icrbe;	/* II CRB Register D */
+	int i;
+	int num_errors = 0;	/* Num of errors handled */
+	ioerror_t ioerror;
+
+	nasid = hubdev_info->hdi_nasid;
+
+	/*
+	 * XXX - Add locking for any recovery actions
+	 */
+	/*
+	 * Scan through all CRBs in the Hub, and handle the errors
+	 * in any of the CRBs marked.
+	 */
+	for (i = 0; i < IIO_NUM_CRBS; i++) {
+		/* Check this crb entry to see if it is in error. */
+		icrbb.ii_icrb0_b_regval = REMOTE_HUB_L(nasid, IIO_ICRB_B(i));
+
+		if (icrbb.b_mark == 0) {
+			continue;
+		}
+
+		icrba.ii_icrb0_a_regval = REMOTE_HUB_L(nasid, IIO_ICRB_A(i));
+
+		IOERROR_INIT(&ioerror);
+
+		/* read other CRB error registers. */
+		icrbc.ii_icrb0_c_regval = REMOTE_HUB_L(nasid, IIO_ICRB_C(i));
+		icrbd.ii_icrb0_d_regval = REMOTE_HUB_L(nasid, IIO_ICRB_D(i));
+		icrbe.ii_icrb0_e_regval = REMOTE_HUB_L(nasid, IIO_ICRB_E(i));
+
+		IOERROR_SETVALUE(&ioerror, errortype, icrbb.b_ecode);
+
+		/* Check if this error is due to BTE operation,
+		 * and handle it separately.
+		 */
+		if (icrbd.d_bteop ||
+		    ((icrbb.b_initiator == IIO_ICRB_INIT_BTE0 ||
+		      icrbb.b_initiator == IIO_ICRB_INIT_BTE1) &&
+		     (icrbb.b_imsgtype == IIO_ICRB_IMSGT_BTE ||
+		      icrbb.b_imsgtype == IIO_ICRB_IMSGT_SN1NET))) {
+
+			int bte_num;
+
+			if (icrbd.d_bteop)
+				bte_num = icrbc.c_btenum;
+			else	/* b_initiator bit 2 gives BTE number */
+				bte_num = (icrbb.b_initiator & 0x4) >> 2;
+
+			hubiio_crb_free(hubdev_info, i);
+
+			bte_crb_error_handler(nasid_to_cnodeid(nasid), bte_num,
+					      i, &ioerror, icrbd.d_bteop);
+			num_errors++;
+			continue;
+		}
+	}
+}
+
+/*
+ * Function	: hub_error_init
+ * Purpose	: initialize the error handling requirements for a given hub.
+ * Parameters	: cnode, the compact nodeid.
+ * Assumptions	: Called only once per hub, either by a local cpu. Or by a
+ *			remote cpu, when this hub is headless.(cpuless)
+ * Returns	: None
+ */
+void hub_error_init(struct hubdev_info *hubdev_info)
+{
+
+	if (request_irq(SGI_II_ERROR, hub_eint_handler, IRQF_SHARED,
+			"SN_hub_error", hubdev_info)) {
+		printk(KERN_ERR "hub_error_init: Failed to request_irq for 0x%p\n",
+		    hubdev_info);
+		return;
+	}
+	irq_set_handler(SGI_II_ERROR, handle_level_irq);
+	sn_set_err_irq_affinity(SGI_II_ERROR);
+}
+
+
+/*
+ * Function	: ice_error_init
+ * Purpose	: initialize the error handling requirements for a given tio.
+ * Parameters	: cnode, the compact nodeid.
+ * Assumptions	: Called only once per tio.
+ * Returns	: None
+ */
+void ice_error_init(struct hubdev_info *hubdev_info)
+{
+
+        if (request_irq
+            (SGI_TIO_ERROR, (void *)hub_eint_handler, IRQF_SHARED, "SN_TIO_error",
+             (void *)hubdev_info)) {
+                printk("ice_error_init: request_irq() error hubdev_info 0x%p\n",
+                       hubdev_info);
+		return;
+	}
+	irq_set_handler(SGI_TIO_ERROR, handle_level_irq);
+	sn_set_err_irq_affinity(SGI_TIO_ERROR);
+}
+
diff --git a/kernel/arch/ia64/sn/kernel/idle.c b/kernel/arch/ia64/sn/kernel/idle.c
new file mode 100644
index 000000000..49d178f02
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/idle.c
@@ -0,0 +1,30 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2001-2004 Silicon Graphics, Inc.  All rights reserved.
+ */
+
+#include <asm/sn/leds.h>
+
+void snidle(int state)
+{
+	if (state) {
+		if (pda->idle_flag == 0) {
+			/* 
+			 * Turn the activity LED off.
+			 */
+			set_led_bits(0, LED_CPU_ACTIVITY);
+		}
+
+		pda->idle_flag = 1;
+	} else {
+		/* 
+		 * Turn the activity LED on.
+		 */
+		set_led_bits(LED_CPU_ACTIVITY, LED_CPU_ACTIVITY);
+
+		pda->idle_flag = 0;
+	}
+}
diff --git a/kernel/arch/ia64/sn/kernel/io_acpi_init.c b/kernel/arch/ia64/sn/kernel/io_acpi_init.c
new file mode 100644
index 000000000..0640739cc
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/io_acpi_init.c
@@ -0,0 +1,510 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <asm/sn/types.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/sn_sal.h>
+#include "xtalk/hubdev.h"
+#include <linux/acpi.h>
+#include <linux/slab.h>
+#include <linux/export.h>
+
+
+/*
+ * The code in this file will only be executed when running with
+ * a PROM that has ACPI IO support. (i.e., SN_ACPI_BASE_SUPPORT() == 1)
+ */
+
+
+/*
+ * This value must match the UUID the PROM uses
+ * (io/acpi/defblk.c) when building a vendor descriptor.
+ */
+struct acpi_vendor_uuid sn_uuid = {
+	.subtype = 0,
+	.data	= { 0x2c, 0xc6, 0xa6, 0xfe, 0x9c, 0x44, 0xda, 0x11,
+		    0xa2, 0x7c, 0x08, 0x00, 0x69, 0x13, 0xea, 0x51 },
+};
+
+struct sn_pcidev_match {
+	u8 bus;
+	unsigned int devfn;
+	acpi_handle handle;
+};
+
+/*
+ * Perform the early IO init in PROM.
+ */
+static long
+sal_ioif_init(u64 *result)
+{
+	struct ia64_sal_retval isrv = {0,0,0,0};
+
+	SAL_CALL_NOLOCK(isrv,
+			SN_SAL_IOIF_INIT, 0, 0, 0, 0, 0, 0, 0);
+	*result = isrv.v0;
+	return isrv.status;
+}
+
+/*
+ * sn_acpi_hubdev_init() - This function is called by acpi_ns_get_device_callback()
+ *			   for all SGIHUB and SGITIO acpi devices defined in the
+ *			   DSDT. It obtains the hubdev_info pointer from the
+ *			   ACPI vendor resource, which the PROM setup, and sets up the
+ *			   hubdev_info in the pda.
+ */
+
+static acpi_status __init
+sn_acpi_hubdev_init(acpi_handle handle, u32 depth, void *context, void **ret)
+{
+	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	u64 addr;
+	struct hubdev_info *hubdev;
+	struct hubdev_info *hubdev_ptr;
+	int i;
+	u64 nasid;
+	struct acpi_resource *resource;
+	acpi_status status;
+	struct acpi_resource_vendor_typed *vendor;
+	extern void sn_common_hubdev_init(struct hubdev_info *);
+
+	status = acpi_get_vendor_resource(handle, METHOD_NAME__CRS,
+					  &sn_uuid, &buffer);
+	if (ACPI_FAILURE(status)) {
+		acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR
+		       "sn_acpi_hubdev_init: acpi_get_vendor_resource() "
+		       "(0x%x) failed for: %s\n", status,
+			(char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		return AE_OK;		/* Continue walking namespace */
+	}
+
+	resource = buffer.pointer;
+	vendor = &resource->data.vendor_typed;
+	if ((vendor->byte_length - sizeof(struct acpi_vendor_uuid)) !=
+	    sizeof(struct hubdev_info *)) {
+		acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR
+		       "sn_acpi_hubdev_init: Invalid vendor data length: "
+		       "%d for: %s\n",
+			vendor->byte_length, (char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		goto exit;
+	}
+
+	memcpy(&addr, vendor->byte_data, sizeof(struct hubdev_info *));
+	hubdev_ptr = __va((struct hubdev_info *) addr);
+
+	nasid = hubdev_ptr->hdi_nasid;
+	i = nasid_to_cnodeid(nasid);
+	hubdev = (struct hubdev_info *)(NODEPDA(i)->pdinfo);
+	*hubdev = *hubdev_ptr;
+	sn_common_hubdev_init(hubdev);
+
+exit:
+	kfree(buffer.pointer);
+	return AE_OK;		/* Continue walking namespace */
+}
+
+/*
+ * sn_get_bussoft_ptr() - The pcibus_bussoft pointer is found in
+ *			  the ACPI Vendor resource for this bus.
+ */
+static struct pcibus_bussoft *
+sn_get_bussoft_ptr(struct pci_bus *bus)
+{
+	u64 addr;
+	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	acpi_handle handle;
+	struct pcibus_bussoft *prom_bussoft_ptr;
+	struct acpi_resource *resource;
+	acpi_status status;
+	struct acpi_resource_vendor_typed *vendor;
+
+
+	handle = acpi_device_handle(PCI_CONTROLLER(bus)->companion);
+	status = acpi_get_vendor_resource(handle, METHOD_NAME__CRS,
+					  &sn_uuid, &buffer);
+	if (ACPI_FAILURE(status)) {
+		acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR "%s: "
+		       "acpi_get_vendor_resource() failed (0x%x) for: %s\n",
+		       __func__, status, (char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		return NULL;
+	}
+	resource = buffer.pointer;
+	vendor = &resource->data.vendor_typed;
+
+	if ((vendor->byte_length - sizeof(struct acpi_vendor_uuid)) !=
+	     sizeof(struct pcibus_bussoft *)) {
+		printk(KERN_ERR
+		       "%s: Invalid vendor data length %d\n",
+			__func__, vendor->byte_length);
+		kfree(buffer.pointer);
+		return NULL;
+	}
+	memcpy(&addr, vendor->byte_data, sizeof(struct pcibus_bussoft *));
+	prom_bussoft_ptr = __va((struct pcibus_bussoft *) addr);
+	kfree(buffer.pointer);
+
+	return prom_bussoft_ptr;
+}
+
+/*
+ * sn_extract_device_info - Extract the pcidev_info and the sn_irq_info
+ *			    pointers from the vendor resource using the
+ *			    provided acpi handle, and copy the structures
+ *			    into the argument buffers.
+ */
+static int
+sn_extract_device_info(acpi_handle handle, struct pcidev_info **pcidev_info,
+		    struct sn_irq_info **sn_irq_info)
+{
+	u64 addr;
+	struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+	struct sn_irq_info *irq_info, *irq_info_prom;
+	struct pcidev_info *pcidev_ptr, *pcidev_prom_ptr;
+	struct acpi_resource *resource;
+	int ret = 0;
+	acpi_status status;
+	struct acpi_resource_vendor_typed *vendor;
+
+	/*
+	 * The pointer to this device's pcidev_info structure in
+	 * the PROM, is in the vendor resource.
+	 */
+	status = acpi_get_vendor_resource(handle, METHOD_NAME__CRS,
+					  &sn_uuid, &buffer);
+	if (ACPI_FAILURE(status)) {
+		acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR
+		       "%s: acpi_get_vendor_resource() failed (0x%x) for: %s\n",
+			__func__, status, (char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		return 1;
+	}
+
+	resource = buffer.pointer;
+	vendor = &resource->data.vendor_typed;
+	if ((vendor->byte_length - sizeof(struct acpi_vendor_uuid)) !=
+	    sizeof(struct pci_devdev_info *)) {
+		acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR
+		       "%s: Invalid vendor data length: %d for: %s\n",
+			 __func__, vendor->byte_length,
+			(char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		ret = 1;
+		goto exit;
+	}
+
+	pcidev_ptr = kzalloc(sizeof(struct pcidev_info), GFP_KERNEL);
+	if (!pcidev_ptr)
+		panic("%s: Unable to alloc memory for pcidev_info", __func__);
+
+	memcpy(&addr, vendor->byte_data, sizeof(struct pcidev_info *));
+	pcidev_prom_ptr = __va(addr);
+	memcpy(pcidev_ptr, pcidev_prom_ptr, sizeof(struct pcidev_info));
+
+	/* Get the IRQ info */
+	irq_info = kzalloc(sizeof(struct sn_irq_info), GFP_KERNEL);
+	if (!irq_info)
+		 panic("%s: Unable to alloc memory for sn_irq_info", __func__);
+
+	if (pcidev_ptr->pdi_sn_irq_info) {
+		irq_info_prom = __va(pcidev_ptr->pdi_sn_irq_info);
+		memcpy(irq_info, irq_info_prom, sizeof(struct sn_irq_info));
+	}
+
+	*pcidev_info = pcidev_ptr;
+	*sn_irq_info = irq_info;
+
+exit:
+	kfree(buffer.pointer);
+	return ret;
+}
+
+static unsigned int
+get_host_devfn(acpi_handle device_handle, acpi_handle rootbus_handle)
+{
+	unsigned long long adr;
+	acpi_handle child;
+	unsigned int devfn;
+	int function;
+	acpi_handle parent;
+	int slot;
+	acpi_status status;
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+
+	acpi_get_name(device_handle, ACPI_FULL_PATHNAME, &name_buffer);
+
+	/*
+	 * Do an upward search to find the root bus device, and
+	 * obtain the host devfn from the previous child device.
+	 */
+	child = device_handle;
+	while (child) {
+		status = acpi_get_parent(child, &parent);
+		if (ACPI_FAILURE(status)) {
+			printk(KERN_ERR "%s: acpi_get_parent() failed "
+			       "(0x%x) for: %s\n", __func__, status,
+				(char *)name_buffer.pointer);
+			panic("%s: Unable to find host devfn\n", __func__);
+		}
+		if (parent == rootbus_handle)
+			break;
+		child = parent;
+	}
+	if (!child) {
+		printk(KERN_ERR "%s: Unable to find root bus for: %s\n",
+		       __func__, (char *)name_buffer.pointer);
+		BUG();
+	}
+
+	status = acpi_evaluate_integer(child, METHOD_NAME__ADR, NULL, &adr);
+	if (ACPI_FAILURE(status)) {
+		printk(KERN_ERR "%s: Unable to get _ADR (0x%x) for: %s\n",
+		       __func__, status, (char *)name_buffer.pointer);
+		panic("%s: Unable to find host devfn\n", __func__);
+	}
+
+	kfree(name_buffer.pointer);
+
+	slot = (adr >> 16) & 0xffff;
+	function = adr & 0xffff;
+	devfn = PCI_DEVFN(slot, function);
+	return devfn;
+}
+
+/*
+ * find_matching_device - Callback routine to find the ACPI device
+ *			  that matches up with our pci_dev device.
+ *			  Matching is done on bus number and devfn.
+ *			  To find the bus number for a particular
+ *			  ACPI device, we must look at the _BBN method
+ *			  of its parent.
+ */
+static acpi_status
+find_matching_device(acpi_handle handle, u32 lvl, void *context, void **rv)
+{
+	unsigned long long bbn = -1;
+	unsigned long long adr;
+	acpi_handle parent = NULL;
+	acpi_status status;
+	unsigned int devfn;
+	int function;
+	int slot;
+	struct sn_pcidev_match *info = context;
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+
+        status = acpi_evaluate_integer(handle, METHOD_NAME__ADR, NULL,
+                                       &adr);
+        if (ACPI_SUCCESS(status)) {
+		status = acpi_get_parent(handle, &parent);
+		if (ACPI_FAILURE(status)) {
+			acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+			printk(KERN_ERR
+			       "%s: acpi_get_parent() failed (0x%x) for: %s\n",
+				__func__, status, (char *)name_buffer.pointer);
+			kfree(name_buffer.pointer);
+			return AE_OK;
+		}
+		status = acpi_evaluate_integer(parent, METHOD_NAME__BBN,
+					       NULL, &bbn);
+		if (ACPI_FAILURE(status)) {
+			acpi_get_name(handle, ACPI_FULL_PATHNAME, &name_buffer);
+			printk(KERN_ERR
+			  "%s: Failed to find _BBN in parent of: %s\n",
+					__func__, (char *)name_buffer.pointer);
+			kfree(name_buffer.pointer);
+			return AE_OK;
+		}
+
+                slot = (adr >> 16) & 0xffff;
+                function = adr & 0xffff;
+                devfn = PCI_DEVFN(slot, function);
+                if ((info->devfn == devfn) && (info->bus == bbn)) {
+			/* We have a match! */
+			info->handle = handle;
+			return 1;
+		}
+	}
+	return AE_OK;
+}
+
+/*
+ * sn_acpi_get_pcidev_info - Search ACPI namespace for the acpi
+ *			     device matching the specified pci_dev,
+ *			     and return the pcidev info and irq info.
+ */
+int
+sn_acpi_get_pcidev_info(struct pci_dev *dev, struct pcidev_info **pcidev_info,
+			struct sn_irq_info **sn_irq_info)
+{
+	unsigned int host_devfn;
+	struct sn_pcidev_match pcidev_match;
+	acpi_handle rootbus_handle;
+	unsigned long long segment;
+	acpi_status status;
+	struct acpi_buffer name_buffer = { ACPI_ALLOCATE_BUFFER, NULL };
+
+	rootbus_handle = acpi_device_handle(PCI_CONTROLLER(dev)->companion);
+        status = acpi_evaluate_integer(rootbus_handle, METHOD_NAME__SEG, NULL,
+                                       &segment);
+        if (ACPI_SUCCESS(status)) {
+		if (segment != pci_domain_nr(dev)) {
+			acpi_get_name(rootbus_handle, ACPI_FULL_PATHNAME,
+				&name_buffer);
+			printk(KERN_ERR
+			       "%s: Segment number mismatch, 0x%llx vs 0x%x for: %s\n",
+			       __func__, segment, pci_domain_nr(dev),
+			       (char *)name_buffer.pointer);
+			kfree(name_buffer.pointer);
+			return 1;
+		}
+	} else {
+		acpi_get_name(rootbus_handle, ACPI_FULL_PATHNAME, &name_buffer);
+		printk(KERN_ERR "%s: Unable to get __SEG from: %s\n",
+		       __func__, (char *)name_buffer.pointer);
+		kfree(name_buffer.pointer);
+		return 1;
+	}
+
+	/*
+	 * We want to search all devices in this segment/domain
+	 * of the ACPI namespace for the matching ACPI device,
+	 * which holds the pcidev_info pointer in its vendor resource.
+	 */
+	pcidev_match.bus = dev->bus->number;
+	pcidev_match.devfn = dev->devfn;
+	pcidev_match.handle = NULL;
+
+	acpi_walk_namespace(ACPI_TYPE_DEVICE, rootbus_handle, ACPI_UINT32_MAX,
+			    find_matching_device, NULL, &pcidev_match, NULL);
+
+	if (!pcidev_match.handle) {
+		printk(KERN_ERR
+		       "%s: Could not find matching ACPI device for %s.\n",
+		       __func__, pci_name(dev));
+		return 1;
+	}
+
+	if (sn_extract_device_info(pcidev_match.handle, pcidev_info, sn_irq_info))
+		return 1;
+
+	/* Build up the pcidev_info.pdi_slot_host_handle */
+	host_devfn = get_host_devfn(pcidev_match.handle, rootbus_handle);
+	(*pcidev_info)->pdi_slot_host_handle =
+			((unsigned long) pci_domain_nr(dev) << 40) |
+					/* bus == 0 */
+					host_devfn;
+	return 0;
+}
+
+/*
+ * sn_acpi_slot_fixup - Obtain the pcidev_info and sn_irq_info.
+ *			Perform any SN specific slot fixup.
+ *			At present there does not appear to be
+ *			any generic way to handle a ROM image
+ *			that has been shadowed by the PROM, so
+ *			we pass a pointer to it	within the
+ *			pcidev_info structure.
+ */
+
+void
+sn_acpi_slot_fixup(struct pci_dev *dev)
+{
+	void __iomem *addr;
+	struct pcidev_info *pcidev_info = NULL;
+	struct sn_irq_info *sn_irq_info = NULL;
+	size_t image_size, size;
+
+	if (sn_acpi_get_pcidev_info(dev, &pcidev_info, &sn_irq_info)) {
+		panic("%s:  Failure obtaining pcidev_info for %s\n",
+		      __func__, pci_name(dev));
+	}
+
+	if (pcidev_info->pdi_pio_mapped_addr[PCI_ROM_RESOURCE]) {
+		/*
+		 * A valid ROM image exists and has been shadowed by the
+		 * PROM. Setup the pci_dev ROM resource with the address
+		 * of the shadowed copy, and the actual length of the ROM image.
+		 */
+		size = pci_resource_len(dev, PCI_ROM_RESOURCE);
+		addr = ioremap(pcidev_info->pdi_pio_mapped_addr[PCI_ROM_RESOURCE],
+			       size);
+		image_size = pci_get_rom_size(dev, addr, size);
+		dev->resource[PCI_ROM_RESOURCE].start = (unsigned long) addr;
+		dev->resource[PCI_ROM_RESOURCE].end =
+					(unsigned long) addr + image_size - 1;
+		dev->resource[PCI_ROM_RESOURCE].flags |= IORESOURCE_ROM_BIOS_COPY;
+	}
+	sn_pci_fixup_slot(dev, pcidev_info, sn_irq_info);
+}
+
+EXPORT_SYMBOL(sn_acpi_slot_fixup);
+
+
+/*
+ * sn_acpi_bus_fixup -  Perform SN specific setup of software structs
+ *			(pcibus_bussoft, pcidev_info) and hardware
+ *			registers, for the specified bus and devices under it.
+ */
+void
+sn_acpi_bus_fixup(struct pci_bus *bus)
+{
+	struct pci_dev *pci_dev = NULL;
+	struct pcibus_bussoft *prom_bussoft_ptr;
+
+	if (!bus->parent) {	/* If root bus */
+		prom_bussoft_ptr = sn_get_bussoft_ptr(bus);
+		if (prom_bussoft_ptr == NULL) {
+			printk(KERN_ERR
+			       "%s: 0x%04x:0x%02x Unable to "
+			       "obtain prom_bussoft_ptr\n",
+			       __func__, pci_domain_nr(bus), bus->number);
+			return;
+		}
+		sn_common_bus_fixup(bus, prom_bussoft_ptr);
+	}
+	list_for_each_entry(pci_dev, &bus->devices, bus_list) {
+		sn_acpi_slot_fixup(pci_dev);
+	}
+}
+
+/*
+ * sn_io_acpi_init - PROM has ACPI support for IO, defining at a minimum the
+ *		     nodes and root buses in the DSDT. As a result, bus scanning
+ *		     will be initiated by the Linux ACPI code.
+ */
+
+void __init
+sn_io_acpi_init(void)
+{
+	u64 result;
+	long status;
+
+	/* SN Altix does not follow the IOSAPIC IRQ routing model */
+	acpi_irq_model = ACPI_IRQ_MODEL_PLATFORM;
+
+	/* Setup hubdev_info for all SGIHUB/SGITIO devices */
+	acpi_get_devices("SGIHUB", sn_acpi_hubdev_init, NULL, NULL);
+	acpi_get_devices("SGITIO", sn_acpi_hubdev_init, NULL, NULL);
+
+	status = sal_ioif_init(&result);
+	if (status || result)
+		panic("sal_ioif_init failed: [%lx] %s\n",
+		      status, ia64_sal_strerror(status));
+}
diff --git a/kernel/arch/ia64/sn/kernel/io_common.c b/kernel/arch/ia64/sn/kernel/io_common.c
new file mode 100644
index 000000000..11f227557
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/io_common.c
@@ -0,0 +1,564 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/bootmem.h>
+#include <linux/export.h>
+#include <linux/slab.h>
+#include <asm/sn/types.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/sn_feature_sets.h>
+#include <asm/sn/geo.h>
+#include <asm/sn/io.h>
+#include <asm/sn/l1.h>
+#include <asm/sn/module.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/simulator.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/tioca_provider.h>
+#include <asm/sn/tioce_provider.h>
+#include "xtalk/hubdev.h"
+#include "xtalk/xwidgetdev.h"
+#include <linux/acpi.h>
+#include <asm/sn/sn2/sn_hwperf.h>
+#include <asm/sn/acpi.h>
+
+extern void sn_init_cpei_timer(void);
+extern void register_sn_procfs(void);
+extern void sn_io_acpi_init(void);
+extern void sn_io_init(void);
+
+
+static struct list_head sn_sysdata_list;
+
+/* sysdata list struct */
+struct sysdata_el {
+	struct list_head entry;
+	void *sysdata;
+};
+
+int sn_ioif_inited;		/* SN I/O infrastructure initialized? */
+
+int sn_acpi_rev;		/* SN ACPI revision */
+EXPORT_SYMBOL_GPL(sn_acpi_rev);
+
+struct sn_pcibus_provider *sn_pci_provider[PCIIO_ASIC_MAX_TYPES];	/* indexed by asic type */
+
+/*
+ * Hooks and struct for unsupported pci providers
+ */
+
+static dma_addr_t
+sn_default_pci_map(struct pci_dev *pdev, unsigned long paddr, size_t size, int type)
+{
+	return 0;
+}
+
+static void
+sn_default_pci_unmap(struct pci_dev *pdev, dma_addr_t addr, int direction)
+{
+	return;
+}
+
+static void *
+sn_default_pci_bus_fixup(struct pcibus_bussoft *soft, struct pci_controller *controller)
+{
+	return NULL;
+}
+
+static struct sn_pcibus_provider sn_pci_default_provider = {
+	.dma_map = sn_default_pci_map,
+	.dma_map_consistent = sn_default_pci_map,
+	.dma_unmap = sn_default_pci_unmap,
+	.bus_fixup = sn_default_pci_bus_fixup,
+};
+
+/*
+ * Retrieve the DMA Flush List given nasid, widget, and device.
+ * This list is needed to implement the WAR - Flush DMA data on PIO Reads.
+ */
+static inline u64
+sal_get_device_dmaflush_list(u64 nasid, u64 widget_num, u64 device_num,
+			     u64 address)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_GET_DEVICE_DMAFLUSH_LIST,
+			(u64) nasid, (u64) widget_num,
+			(u64) device_num, (u64) address, 0, 0, 0);
+	return ret_stuff.status;
+}
+
+/*
+ * sn_pcidev_info_get() - Retrieve the pcidev_info struct for the specified
+ *			  device.
+ */
+inline struct pcidev_info *
+sn_pcidev_info_get(struct pci_dev *dev)
+{
+	struct pcidev_info *pcidev;
+
+	list_for_each_entry(pcidev,
+			    &(SN_PLATFORM_DATA(dev)->pcidev_info), pdi_list) {
+		if (pcidev->pdi_linux_pcidev == dev)
+			return pcidev;
+	}
+	return NULL;
+}
+
+/* Older PROM flush WAR
+ *
+ * 01/16/06 -- This war will be in place until a new official PROM is released.
+ * Additionally note that the struct sn_flush_device_war also has to be
+ * removed from arch/ia64/sn/include/xtalk/hubdev.h
+ */
+
+static s64 sn_device_fixup_war(u64 nasid, u64 widget, int device,
+			       struct sn_flush_device_common *common)
+{
+	struct sn_flush_device_war *war_list;
+	struct sn_flush_device_war *dev_entry;
+	struct ia64_sal_retval isrv = {0,0,0,0};
+
+	printk_once(KERN_WARNING
+		"PROM version < 4.50 -- implementing old PROM flush WAR\n");
+
+	war_list = kzalloc(DEV_PER_WIDGET * sizeof(*war_list), GFP_KERNEL);
+	BUG_ON(!war_list);
+
+	SAL_CALL_NOLOCK(isrv, SN_SAL_IOIF_GET_WIDGET_DMAFLUSH_LIST,
+			nasid, widget, __pa(war_list), 0, 0, 0 ,0);
+	if (isrv.status)
+		panic("sn_device_fixup_war failed: %s\n",
+		      ia64_sal_strerror(isrv.status));
+
+	dev_entry = war_list + device;
+	memcpy(common,dev_entry, sizeof(*common));
+	kfree(war_list);
+
+	return isrv.status;
+}
+
+/*
+ * sn_common_hubdev_init() - This routine is called to initialize the HUB data
+ *			     structure for each node in the system.
+ */
+void __init
+sn_common_hubdev_init(struct hubdev_info *hubdev)
+{
+
+	struct sn_flush_device_kernel *sn_flush_device_kernel;
+	struct sn_flush_device_kernel *dev_entry;
+	s64 status;
+	int widget, device, size;
+
+	/* Attach the error interrupt handlers */
+	if (hubdev->hdi_nasid & 1)	/* If TIO */
+		ice_error_init(hubdev);
+	else
+		hub_error_init(hubdev);
+
+	for (widget = 0; widget <= HUB_WIDGET_ID_MAX; widget++)
+		hubdev->hdi_xwidget_info[widget].xwi_hubinfo = hubdev;
+
+	if (!hubdev->hdi_flush_nasid_list.widget_p)
+		return;
+
+	size = (HUB_WIDGET_ID_MAX + 1) *
+		sizeof(struct sn_flush_device_kernel *);
+	hubdev->hdi_flush_nasid_list.widget_p =
+		kzalloc(size, GFP_KERNEL);
+	BUG_ON(!hubdev->hdi_flush_nasid_list.widget_p);
+
+	for (widget = 0; widget <= HUB_WIDGET_ID_MAX; widget++) {
+		size = DEV_PER_WIDGET *
+			sizeof(struct sn_flush_device_kernel);
+		sn_flush_device_kernel = kzalloc(size, GFP_KERNEL);
+		BUG_ON(!sn_flush_device_kernel);
+
+		dev_entry = sn_flush_device_kernel;
+		for (device = 0; device < DEV_PER_WIDGET;
+		     device++, dev_entry++) {
+			size = sizeof(struct sn_flush_device_common);
+			dev_entry->common = kzalloc(size, GFP_KERNEL);
+			BUG_ON(!dev_entry->common);
+			if (sn_prom_feature_available(PRF_DEVICE_FLUSH_LIST))
+				status = sal_get_device_dmaflush_list(
+					     hubdev->hdi_nasid, widget, device,
+					     (u64)(dev_entry->common));
+			else
+				status = sn_device_fixup_war(hubdev->hdi_nasid,
+							     widget, device,
+							     dev_entry->common);
+			if (status != SALRET_OK)
+				panic("SAL call failed: %s\n",
+				      ia64_sal_strerror(status));
+
+			spin_lock_init(&dev_entry->sfdl_flush_lock);
+		}
+
+		if (sn_flush_device_kernel)
+			hubdev->hdi_flush_nasid_list.widget_p[widget] =
+							 sn_flush_device_kernel;
+	}
+}
+
+void sn_pci_unfixup_slot(struct pci_dev *dev)
+{
+	struct pci_dev *host_pci_dev = SN_PCIDEV_INFO(dev)->host_pci_dev;
+
+	sn_irq_unfixup(dev);
+	pci_dev_put(host_pci_dev);
+	pci_dev_put(dev);
+}
+
+/*
+ * sn_pci_fixup_slot()
+ */
+void sn_pci_fixup_slot(struct pci_dev *dev, struct pcidev_info *pcidev_info,
+		       struct sn_irq_info *sn_irq_info)
+{
+	int segment = pci_domain_nr(dev->bus);
+	struct pcibus_bussoft *bs;
+	struct pci_dev *host_pci_dev;
+	unsigned int bus_no, devfn;
+
+	pci_dev_get(dev); /* for the sysdata pointer */
+
+	/* Add pcidev_info to list in pci_controller.platform_data */
+	list_add_tail(&pcidev_info->pdi_list,
+		      &(SN_PLATFORM_DATA(dev->bus)->pcidev_info));
+	/*
+	 * Using the PROMs values for the PCI host bus, get the Linux
+	 * PCI host_pci_dev struct and set up host bus linkages
+ 	 */
+
+	bus_no = (pcidev_info->pdi_slot_host_handle >> 32) & 0xff;
+	devfn = pcidev_info->pdi_slot_host_handle & 0xffffffff;
+	host_pci_dev = pci_get_domain_bus_and_slot(segment, bus_no, devfn);
+
+	pcidev_info->host_pci_dev = host_pci_dev;
+	pcidev_info->pdi_linux_pcidev = dev;
+	pcidev_info->pdi_host_pcidev_info = SN_PCIDEV_INFO(host_pci_dev);
+	bs = SN_PCIBUS_BUSSOFT(dev->bus);
+	pcidev_info->pdi_pcibus_info = bs;
+
+	if (bs && bs->bs_asic_type < PCIIO_ASIC_MAX_TYPES) {
+		SN_PCIDEV_BUSPROVIDER(dev) = sn_pci_provider[bs->bs_asic_type];
+	} else {
+		SN_PCIDEV_BUSPROVIDER(dev) = &sn_pci_default_provider;
+	}
+
+	/* Only set up IRQ stuff if this device has a host bus context */
+	if (bs && sn_irq_info->irq_irq) {
+		pcidev_info->pdi_sn_irq_info = sn_irq_info;
+		dev->irq = pcidev_info->pdi_sn_irq_info->irq_irq;
+		sn_irq_fixup(dev, sn_irq_info);
+	} else {
+		pcidev_info->pdi_sn_irq_info = NULL;
+		kfree(sn_irq_info);
+	}
+}
+
+/*
+ * sn_common_bus_fixup - Perform platform specific bus fixup.
+ *			 Execute the ASIC specific fixup routine
+ *			 for this bus.
+ */
+void
+sn_common_bus_fixup(struct pci_bus *bus,
+		    struct pcibus_bussoft *prom_bussoft_ptr)
+{
+	int cnode;
+	struct pci_controller *controller;
+	struct hubdev_info *hubdev_info;
+	int nasid;
+	void *provider_soft;
+	struct sn_pcibus_provider *provider;
+	struct sn_platform_data *sn_platform_data;
+
+	controller = PCI_CONTROLLER(bus);
+	/*
+	 * Per-provider fixup.  Copies the bus soft structure from prom
+	 * to local area and links SN_PCIBUS_BUSSOFT().
+	 */
+
+	if (prom_bussoft_ptr->bs_asic_type >= PCIIO_ASIC_MAX_TYPES) {
+		printk(KERN_WARNING "sn_common_bus_fixup: Unsupported asic type, %d",
+		       prom_bussoft_ptr->bs_asic_type);
+		return;
+	}
+
+	if (prom_bussoft_ptr->bs_asic_type == PCIIO_ASIC_TYPE_PPB)
+		return;	/* no further fixup necessary */
+
+	provider = sn_pci_provider[prom_bussoft_ptr->bs_asic_type];
+	if (provider == NULL)
+		panic("sn_common_bus_fixup: No provider registered for this asic type, %d",
+		      prom_bussoft_ptr->bs_asic_type);
+
+	if (provider->bus_fixup)
+		provider_soft = (*provider->bus_fixup) (prom_bussoft_ptr,
+				 controller);
+	else
+		provider_soft = NULL;
+
+	/*
+	 * Generic bus fixup goes here.  Don't reference prom_bussoft_ptr
+	 * after this point.
+	 */
+	controller->platform_data = kzalloc(sizeof(struct sn_platform_data),
+					    GFP_KERNEL);
+	BUG_ON(controller->platform_data == NULL);
+	sn_platform_data =
+			(struct sn_platform_data *) controller->platform_data;
+	sn_platform_data->provider_soft = provider_soft;
+	INIT_LIST_HEAD(&((struct sn_platform_data *)
+			 controller->platform_data)->pcidev_info);
+	nasid = NASID_GET(SN_PCIBUS_BUSSOFT(bus)->bs_base);
+	cnode = nasid_to_cnodeid(nasid);
+	hubdev_info = (struct hubdev_info *)(NODEPDA(cnode)->pdinfo);
+	SN_PCIBUS_BUSSOFT(bus)->bs_xwidget_info =
+	    &(hubdev_info->hdi_xwidget_info[SN_PCIBUS_BUSSOFT(bus)->bs_xid]);
+
+	/*
+	 * If the node information we obtained during the fixup phase is
+	 * invalid then set controller->node to -1 (undetermined)
+	 */
+	if (controller->node >= num_online_nodes()) {
+		struct pcibus_bussoft *b = SN_PCIBUS_BUSSOFT(bus);
+
+		printk(KERN_WARNING "Device ASIC=%u XID=%u PBUSNUM=%u "
+		       "L_IO=%llx L_MEM=%llx BASE=%llx\n",
+		       b->bs_asic_type, b->bs_xid, b->bs_persist_busnum,
+		       b->bs_legacy_io, b->bs_legacy_mem, b->bs_base);
+		printk(KERN_WARNING "on node %d but only %d nodes online."
+		       "Association set to undetermined.\n",
+		       controller->node, num_online_nodes());
+		controller->node = -1;
+	}
+}
+
+void sn_bus_store_sysdata(struct pci_dev *dev)
+{
+	struct sysdata_el *element;
+
+	element = kzalloc(sizeof(struct sysdata_el), GFP_KERNEL);
+	if (!element) {
+		dev_dbg(&dev->dev, "%s: out of memory!\n", __func__);
+		return;
+	}
+	element->sysdata = SN_PCIDEV_INFO(dev);
+	list_add(&element->entry, &sn_sysdata_list);
+}
+
+void sn_bus_free_sysdata(void)
+{
+	struct sysdata_el *element;
+	struct list_head *list, *safe;
+
+	list_for_each_safe(list, safe, &sn_sysdata_list) {
+		element = list_entry(list, struct sysdata_el, entry);
+		list_del(&element->entry);
+		list_del(&(((struct pcidev_info *)
+			     (element->sysdata))->pdi_list));
+		kfree(element->sysdata);
+		kfree(element);
+	}
+	return;
+}
+
+/*
+ * hubdev_init_node() - Creates the HUB data structure and link them to it's
+ *			own NODE specific data area.
+ */
+void __init hubdev_init_node(nodepda_t * npda, cnodeid_t node)
+{
+	struct hubdev_info *hubdev_info;
+	int size;
+	pg_data_t *pg;
+
+	size = sizeof(struct hubdev_info);
+
+	if (node >= num_online_nodes())	/* Headless/memless IO nodes */
+		pg = NODE_DATA(0);
+	else
+		pg = NODE_DATA(node);
+
+	hubdev_info = (struct hubdev_info *)alloc_bootmem_node(pg, size);
+
+	npda->pdinfo = (void *)hubdev_info;
+}
+
+geoid_t
+cnodeid_get_geoid(cnodeid_t cnode)
+{
+	struct hubdev_info *hubdev;
+
+	hubdev = (struct hubdev_info *)(NODEPDA(cnode)->pdinfo);
+	return hubdev->hdi_geoid;
+}
+
+void sn_generate_path(struct pci_bus *pci_bus, char *address)
+{
+	nasid_t nasid;
+	cnodeid_t cnode;
+	geoid_t geoid;
+	moduleid_t moduleid;
+	u16 bricktype;
+
+	nasid = NASID_GET(SN_PCIBUS_BUSSOFT(pci_bus)->bs_base);
+	cnode = nasid_to_cnodeid(nasid);
+	geoid = cnodeid_get_geoid(cnode);
+	moduleid = geo_module(geoid);
+
+	sprintf(address, "module_%c%c%c%c%.2d",
+		'0'+RACK_GET_CLASS(MODULE_GET_RACK(moduleid)),
+		'0'+RACK_GET_GROUP(MODULE_GET_RACK(moduleid)),
+		'0'+RACK_GET_NUM(MODULE_GET_RACK(moduleid)),
+		MODULE_GET_BTCHAR(moduleid), MODULE_GET_BPOS(moduleid));
+
+	/* Tollhouse requires slot id to be displayed */
+	bricktype = MODULE_GET_BTYPE(moduleid);
+	if ((bricktype == L1_BRICKTYPE_191010) ||
+	    (bricktype == L1_BRICKTYPE_1932))
+			sprintf(address + strlen(address), "^%d",
+						geo_slot(geoid));
+}
+
+void sn_pci_fixup_bus(struct pci_bus *bus)
+{
+
+	if (SN_ACPI_BASE_SUPPORT())
+		sn_acpi_bus_fixup(bus);
+	else
+		sn_bus_fixup(bus);
+}
+
+/*
+ * sn_io_early_init - Perform early IO (and some non-IO) initialization.
+ *		      In particular, setup the sn_pci_provider[] array.
+ *		      This needs to be done prior to any bus scanning
+ *		      (acpi_scan_init()) in the ACPI case, as the SN
+ *		      bus fixup code will reference the array.
+ */
+static int __init
+sn_io_early_init(void)
+{
+	int i;
+
+	if (!ia64_platform_is("sn2") || IS_RUNNING_ON_FAKE_PROM())
+		return 0;
+
+	/* we set the acpi revision to that of the DSDT table OEM rev. */
+	{
+		struct acpi_table_header *header = NULL;
+
+		acpi_get_table(ACPI_SIG_DSDT, 1, &header);
+		BUG_ON(header == NULL);
+		sn_acpi_rev = header->oem_revision;
+	}
+
+	/*
+	 * prime sn_pci_provider[].  Individual provider init routines will
+	 * override their respective default entries.
+	 */
+
+	for (i = 0; i < PCIIO_ASIC_MAX_TYPES; i++)
+		sn_pci_provider[i] = &sn_pci_default_provider;
+
+	pcibr_init_provider();
+	tioca_init_provider();
+	tioce_init_provider();
+
+	/*
+	 * This is needed to avoid bounce limit checks in the blk layer
+	 */
+	ia64_max_iommu_merge_mask = ~PAGE_MASK;
+
+	sn_irq_lh_init();
+	INIT_LIST_HEAD(&sn_sysdata_list);
+	sn_init_cpei_timer();
+
+#ifdef CONFIG_PROC_FS
+	register_sn_procfs();
+#endif
+
+	{
+		struct acpi_table_header *header;
+		(void)acpi_get_table(ACPI_SIG_DSDT, 1, &header);
+		printk(KERN_INFO "ACPI  DSDT OEM Rev 0x%x\n",
+			header->oem_revision);
+	}
+	if (SN_ACPI_BASE_SUPPORT())
+		sn_io_acpi_init();
+	else
+		sn_io_init();
+	return 0;
+}
+
+arch_initcall(sn_io_early_init);
+
+/*
+ * sn_io_late_init() - Perform any final platform specific IO initialization.
+ */
+
+int __init
+sn_io_late_init(void)
+{
+	struct pci_bus *bus;
+	struct pcibus_bussoft *bussoft;
+	cnodeid_t cnode;
+	nasid_t nasid;
+	cnodeid_t near_cnode;
+
+	if (!ia64_platform_is("sn2") || IS_RUNNING_ON_FAKE_PROM())
+		return 0;
+
+	/*
+	 * Setup closest node in pci_controller->node for
+	 * PIC, TIOCP, TIOCE (TIOCA does it during bus fixup using
+	 * info from the PROM).
+	 */
+	bus = NULL;
+	while ((bus = pci_find_next_bus(bus)) != NULL) {
+		bussoft = SN_PCIBUS_BUSSOFT(bus);
+		nasid = NASID_GET(bussoft->bs_base);
+		cnode = nasid_to_cnodeid(nasid);
+		if ((bussoft->bs_asic_type == PCIIO_ASIC_TYPE_TIOCP) ||
+		    (bussoft->bs_asic_type == PCIIO_ASIC_TYPE_TIOCE) ||
+		    (bussoft->bs_asic_type == PCIIO_ASIC_TYPE_PIC)) {
+			/* PCI Bridge: find nearest node with CPUs */
+			int e = sn_hwperf_get_nearest_node(cnode, NULL,
+							   &near_cnode);
+			if (e < 0) {
+				near_cnode = (cnodeid_t)-1; /* use any node */
+				printk(KERN_WARNING "sn_io_late_init: failed "
+				       "to find near node with CPUs for "
+				       "node %d, err=%d\n", cnode, e);
+			}
+			PCI_CONTROLLER(bus)->node = near_cnode;
+		}
+	}
+
+	sn_ioif_inited = 1;	/* SN I/O infrastructure now initialized */
+
+	return 0;
+}
+
+fs_initcall(sn_io_late_init);
+
+EXPORT_SYMBOL(sn_pci_unfixup_slot);
+EXPORT_SYMBOL(sn_bus_store_sysdata);
+EXPORT_SYMBOL(sn_bus_free_sysdata);
+EXPORT_SYMBOL(sn_generate_path);
+
diff --git a/kernel/arch/ia64/sn/kernel/io_init.c b/kernel/arch/ia64/sn/kernel/io_init.c
new file mode 100644
index 000000000..1be65eb07
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/io_init.c
@@ -0,0 +1,321 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992 - 1997, 2000-2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/slab.h>
+#include <linux/export.h>
+#include <asm/sn/types.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/io.h>
+#include <asm/sn/module.h>
+#include <asm/sn/intr.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/sn_sal.h>
+#include "xtalk/hubdev.h"
+
+/*
+ * The code in this file will only be executed when running with
+ * a PROM that does _not_ have base ACPI IO support.
+ * (i.e., SN_ACPI_BASE_SUPPORT() == 0)
+ */
+
+static int max_segment_number;		 /* Default highest segment number */
+static int max_pcibus_number = 255;	/* Default highest pci bus number */
+
+
+/*
+ * Retrieve the hub device info structure for the given nasid.
+ */
+static inline u64 sal_get_hubdev_info(u64 handle, u64 address)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_GET_HUBDEV_INFO,
+			(u64) handle, (u64) address, 0, 0, 0, 0, 0);
+	return ret_stuff.v0;
+}
+
+/*
+ * Retrieve the pci bus information given the bus number.
+ */
+static inline u64 sal_get_pcibus_info(u64 segment, u64 busnum, u64 address)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_GET_PCIBUS_INFO,
+			(u64) segment, (u64) busnum, (u64) address, 0, 0, 0, 0);
+	return ret_stuff.v0;
+}
+
+/*
+ * Retrieve the pci device information given the bus and device|function number.
+ */
+static inline u64
+sal_get_pcidev_info(u64 segment, u64 bus_number, u64 devfn, u64 pci_dev,
+		    u64 sn_irq_info)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_GET_PCIDEV_INFO,
+			(u64) segment, (u64) bus_number, (u64) devfn,
+			(u64) pci_dev,
+			sn_irq_info, 0, 0);
+	return ret_stuff.v0;
+}
+
+
+/*
+ * sn_fixup_ionodes() - This routine initializes the HUB data structure for
+ *			each node in the system. This function is only
+ *			executed when running with a non-ACPI capable PROM.
+ */
+static void __init sn_fixup_ionodes(void)
+{
+
+	struct hubdev_info *hubdev;
+	u64 status;
+	u64 nasid;
+	int i;
+	extern void sn_common_hubdev_init(struct hubdev_info *);
+
+	/*
+	 * Get SGI Specific HUB chipset information.
+	 * Inform Prom that this kernel can support domain bus numbering.
+	 */
+	for (i = 0; i < num_cnodes; i++) {
+		hubdev = (struct hubdev_info *)(NODEPDA(i)->pdinfo);
+		nasid = cnodeid_to_nasid(i);
+		hubdev->max_segment_number = 0xffffffff;
+		hubdev->max_pcibus_number = 0xff;
+		status = sal_get_hubdev_info(nasid, (u64) __pa(hubdev));
+		if (status)
+			continue;
+
+		/* Save the largest Domain and pcibus numbers found. */
+		if (hubdev->max_segment_number) {
+			/*
+			 * Dealing with a Prom that supports segments.
+			 */
+			max_segment_number = hubdev->max_segment_number;
+			max_pcibus_number = hubdev->max_pcibus_number;
+		}
+		sn_common_hubdev_init(hubdev);
+	}
+}
+
+/*
+ * sn_pci_legacy_window_fixup - Setup PCI resources for
+ *				legacy IO and MEM space. This needs to
+ *				be done here, as the PROM does not have
+ *				ACPI support defining the root buses
+ *				and their resources (_CRS),
+ */
+static void
+sn_legacy_pci_window_fixup(struct resource *res,
+		u64 legacy_io, u64 legacy_mem)
+{
+		res[0].name = "legacy_io";
+		res[0].flags = IORESOURCE_IO;
+		res[0].start = legacy_io;
+		res[0].end = res[0].start + 0xffff;
+		res[0].parent = &ioport_resource;
+		res[1].name = "legacy_mem";
+		res[1].flags = IORESOURCE_MEM;
+		res[1].start = legacy_mem;
+		res[1].end = res[1].start + (1024 * 1024) - 1;
+		res[1].parent = &iomem_resource;
+}
+
+/*
+ * sn_io_slot_fixup() -   We are not running with an ACPI capable PROM,
+ *			  and need to convert the pci_dev->resource
+ *			  'start' and 'end' addresses to mapped addresses,
+ *			  and setup the pci_controller->window array entries.
+ */
+void
+sn_io_slot_fixup(struct pci_dev *dev)
+{
+	int idx;
+	unsigned long addr, end, size, start;
+	struct pcidev_info *pcidev_info;
+	struct sn_irq_info *sn_irq_info;
+	int status;
+
+	pcidev_info = kzalloc(sizeof(struct pcidev_info), GFP_KERNEL);
+	if (!pcidev_info)
+		panic("%s: Unable to alloc memory for pcidev_info", __func__);
+
+	sn_irq_info = kzalloc(sizeof(struct sn_irq_info), GFP_KERNEL);
+	if (!sn_irq_info)
+		panic("%s: Unable to alloc memory for sn_irq_info", __func__);
+
+	/* Call to retrieve pci device information needed by kernel. */
+	status = sal_get_pcidev_info((u64) pci_domain_nr(dev),
+		(u64) dev->bus->number,
+		dev->devfn,
+		(u64) __pa(pcidev_info),
+		(u64) __pa(sn_irq_info));
+
+	BUG_ON(status); /* Cannot get platform pci device information */
+
+
+	/* Copy over PIO Mapped Addresses */
+	for (idx = 0; idx <= PCI_ROM_RESOURCE; idx++) {
+
+		if (!pcidev_info->pdi_pio_mapped_addr[idx]) {
+			continue;
+		}
+
+		start = dev->resource[idx].start;
+		end = dev->resource[idx].end;
+		size = end - start;
+		if (size == 0) {
+			continue;
+		}
+		addr = pcidev_info->pdi_pio_mapped_addr[idx];
+		addr = ((addr << 4) >> 4) | __IA64_UNCACHED_OFFSET;
+		dev->resource[idx].start = addr;
+		dev->resource[idx].end = addr + size;
+
+		/*
+		 * if it's already in the device structure, remove it before
+		 * inserting
+		 */
+		if (dev->resource[idx].parent && dev->resource[idx].parent->child)
+			release_resource(&dev->resource[idx]);
+
+		if (dev->resource[idx].flags & IORESOURCE_IO)
+			insert_resource(&ioport_resource, &dev->resource[idx]);
+		else
+			insert_resource(&iomem_resource, &dev->resource[idx]);
+		/*
+		 * If ROM, set the actual ROM image size, and mark as
+		 * shadowed in PROM.
+		 */
+		if (idx == PCI_ROM_RESOURCE) {
+			size_t image_size;
+			void __iomem *rom;
+
+			rom = ioremap(pci_resource_start(dev, PCI_ROM_RESOURCE),
+				      size + 1);
+			image_size = pci_get_rom_size(dev, rom, size + 1);
+			dev->resource[PCI_ROM_RESOURCE].end =
+				dev->resource[PCI_ROM_RESOURCE].start +
+				image_size - 1;
+			dev->resource[PCI_ROM_RESOURCE].flags |=
+						 IORESOURCE_ROM_BIOS_COPY;
+		}
+	}
+
+	sn_pci_fixup_slot(dev, pcidev_info, sn_irq_info);
+}
+
+EXPORT_SYMBOL(sn_io_slot_fixup);
+
+/*
+ * sn_pci_controller_fixup() - This routine sets up a bus's resources
+ *			       consistent with the Linux PCI abstraction layer.
+ */
+static void __init
+sn_pci_controller_fixup(int segment, int busnum, struct pci_bus *bus)
+{
+	s64 status = 0;
+	struct pci_controller *controller;
+	struct pcibus_bussoft *prom_bussoft_ptr;
+	struct resource *res;
+	LIST_HEAD(resources);
+
+ 	status = sal_get_pcibus_info((u64) segment, (u64) busnum,
+ 				     (u64) ia64_tpa(&prom_bussoft_ptr));
+ 	if (status > 0)
+		return;		/*bus # does not exist */
+	prom_bussoft_ptr = __va(prom_bussoft_ptr);
+
+	controller = kzalloc(sizeof(*controller), GFP_KERNEL);
+	BUG_ON(!controller);
+	controller->segment = segment;
+
+	res = kcalloc(2, sizeof(struct resource), GFP_KERNEL);
+	BUG_ON(!res);
+
+	/*
+	 * Temporarily save the prom_bussoft_ptr for use by sn_bus_fixup().
+	 * (platform_data will be overwritten later in sn_common_bus_fixup())
+	 */
+	controller->platform_data = prom_bussoft_ptr;
+
+	sn_legacy_pci_window_fixup(res,
+			prom_bussoft_ptr->bs_legacy_io,
+			prom_bussoft_ptr->bs_legacy_mem);
+	pci_add_resource_offset(&resources,	&res[0],
+			prom_bussoft_ptr->bs_legacy_io);
+	pci_add_resource_offset(&resources,	&res[1],
+			prom_bussoft_ptr->bs_legacy_mem);
+
+	bus = pci_scan_root_bus(NULL, busnum, &pci_root_ops, controller,
+				&resources);
+ 	if (bus == NULL) {
+		kfree(res);
+		kfree(controller);
+		return;
+	}
+	pci_bus_add_devices(bus);
+}
+
+/*
+ * sn_bus_fixup
+ */
+void
+sn_bus_fixup(struct pci_bus *bus)
+{
+	struct pci_dev *pci_dev = NULL;
+	struct pcibus_bussoft *prom_bussoft_ptr;
+
+	if (!bus->parent) {  /* If root bus */
+		prom_bussoft_ptr = PCI_CONTROLLER(bus)->platform_data;
+		if (prom_bussoft_ptr == NULL) {
+			printk(KERN_ERR
+			       "sn_bus_fixup: 0x%04x:0x%02x Unable to "
+			       "obtain prom_bussoft_ptr\n",
+			       pci_domain_nr(bus), bus->number);
+			return;
+		}
+		sn_common_bus_fixup(bus, prom_bussoft_ptr);
+        }
+        list_for_each_entry(pci_dev, &bus->devices, bus_list) {
+                sn_io_slot_fixup(pci_dev);
+        }
+
+}
+
+/*
+ * sn_io_init - PROM does not have ACPI support to define nodes or root buses,
+ *		so we need to do things the hard way, including initiating the
+ *		bus scanning ourselves.
+ */
+
+void __init sn_io_init(void)
+{
+	int i, j;
+
+	sn_fixup_ionodes();
+
+	/* busses are not known yet ... */
+	for (i = 0; i <= max_segment_number; i++)
+		for (j = 0; j <= max_pcibus_number; j++)
+			sn_pci_controller_fixup(i, j, NULL);
+}
diff --git a/kernel/arch/ia64/sn/kernel/iomv.c b/kernel/arch/ia64/sn/kernel/iomv.c
new file mode 100644
index 000000000..c77ebdf98
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/iomv.c
@@ -0,0 +1,82 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2003, 2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/module.h>
+#include <linux/acpi.h>
+#include <asm/io.h>
+#include <asm/delay.h>
+#include <asm/vga.h>
+#include <asm/sn/nodepda.h>
+#include <asm/sn/simulator.h>
+#include <asm/sn/pda.h>
+#include <asm/sn/sn_cpuid.h>
+#include <asm/sn/shub_mmr.h>
+#include <asm/sn/acpi.h>
+
+#define IS_LEGACY_VGA_IOPORT(p) \
+	(((p) >= 0x3b0 && (p) <= 0x3bb) || ((p) >= 0x3c0 && (p) <= 0x3df))
+
+/**
+ * sn_io_addr - convert an in/out port to an i/o address
+ * @port: port to convert
+ *
+ * Legacy in/out instructions are converted to ld/st instructions
+ * on IA64.  This routine will convert a port number into a valid
+ * SN i/o address.  Used by sn_in*() and sn_out*().
+ */
+
+void *sn_io_addr(unsigned long port)
+{
+	if (!IS_RUNNING_ON_SIMULATOR()) {
+		if (IS_LEGACY_VGA_IOPORT(port))
+			return (__ia64_mk_io_addr(port));
+		/* On sn2, legacy I/O ports don't point at anything */
+		if (port < (64 * 1024))
+			return NULL;
+		if (SN_ACPI_BASE_SUPPORT())
+			return (__ia64_mk_io_addr(port));
+		else
+			return ((void *)(port | __IA64_UNCACHED_OFFSET));
+	} else {
+		/* but the simulator uses them... */
+		unsigned long addr;
+
+		/*
+		 * word align port, but need more than 10 bits
+		 * for accessing registers in bedrock local block
+		 * (so we don't do port&0xfff)
+		 */
+		addr = (is_shub2() ? 0xc00000028c000000UL : 0xc0000087cc000000UL) | ((port >> 2) << 12);
+		if ((port >= 0x1f0 && port <= 0x1f7) || port == 0x3f6 || port == 0x3f7)
+			addr |= port;
+		return (void *)addr;
+	}
+}
+
+EXPORT_SYMBOL(sn_io_addr);
+
+/**
+ * __sn_mmiowb - I/O space memory barrier
+ *
+ * See arch/ia64/include/asm/io.h and Documentation/DocBook/deviceiobook.tmpl
+ * for details.
+ *
+ * On SN2, we wait for the PIO_WRITE_STATUS SHub register to clear.
+ * See PV 871084 for details about the WAR about zero value.
+ *
+ */
+void __sn_mmiowb(void)
+{
+	volatile unsigned long *adr = pda->pio_write_status_addr;
+	unsigned long val = pda->pio_write_status_val;
+
+	while ((*adr & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK) != val)
+		cpu_relax();
+}
+
+EXPORT_SYMBOL(__sn_mmiowb);
diff --git a/kernel/arch/ia64/sn/kernel/irq.c b/kernel/arch/ia64/sn/kernel/irq.c
new file mode 100644
index 000000000..85d095154
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/irq.c
@@ -0,0 +1,488 @@
+/*
+ * Platform dependent support for SGI SN
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2000-2008 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/irq.h>
+#include <linux/spinlock.h>
+#include <linux/init.h>
+#include <linux/rculist.h>
+#include <linux/slab.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/arch.h>
+#include <asm/sn/intr.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/shub_mmr.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/sn_feature_sets.h>
+
+static void register_intr_pda(struct sn_irq_info *sn_irq_info);
+static void unregister_intr_pda(struct sn_irq_info *sn_irq_info);
+
+extern int sn_ioif_inited;
+struct list_head **sn_irq_lh;
+static DEFINE_SPINLOCK(sn_irq_info_lock); /* non-IRQ lock */
+
+u64 sn_intr_alloc(nasid_t local_nasid, int local_widget,
+				     struct sn_irq_info *sn_irq_info,
+				     int req_irq, nasid_t req_nasid,
+				     int req_slice)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_INTERRUPT,
+			(u64) SAL_INTR_ALLOC, (u64) local_nasid,
+			(u64) local_widget, __pa(sn_irq_info), (u64) req_irq,
+			(u64) req_nasid, (u64) req_slice);
+
+	return ret_stuff.status;
+}
+
+void sn_intr_free(nasid_t local_nasid, int local_widget,
+				struct sn_irq_info *sn_irq_info)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_INTERRUPT,
+			(u64) SAL_INTR_FREE, (u64) local_nasid,
+			(u64) local_widget, (u64) sn_irq_info->irq_irq,
+			(u64) sn_irq_info->irq_cookie, 0, 0);
+}
+
+u64 sn_intr_redirect(nasid_t local_nasid, int local_widget,
+		      struct sn_irq_info *sn_irq_info,
+		      nasid_t req_nasid, int req_slice)
+{
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_INTERRUPT,
+			(u64) SAL_INTR_REDIRECT, (u64) local_nasid,
+			(u64) local_widget, __pa(sn_irq_info),
+			(u64) req_nasid, (u64) req_slice, 0);
+
+	return ret_stuff.status;
+}
+
+static unsigned int sn_startup_irq(struct irq_data *data)
+{
+	return 0;
+}
+
+static void sn_shutdown_irq(struct irq_data *data)
+{
+}
+
+extern void ia64_mca_register_cpev(int);
+
+static void sn_disable_irq(struct irq_data *data)
+{
+	if (data->irq == local_vector_to_irq(IA64_CPE_VECTOR))
+		ia64_mca_register_cpev(0);
+}
+
+static void sn_enable_irq(struct irq_data *data)
+{
+	if (data->irq == local_vector_to_irq(IA64_CPE_VECTOR))
+		ia64_mca_register_cpev(data->irq);
+}
+
+static void sn_ack_irq(struct irq_data *data)
+{
+	u64 event_occurred, mask;
+	unsigned int irq = data->irq & 0xff;
+
+	event_occurred = HUB_L((u64*)LOCAL_MMR_ADDR(SH_EVENT_OCCURRED));
+	mask = event_occurred & SH_ALL_INT_MASK;
+	HUB_S((u64*)LOCAL_MMR_ADDR(SH_EVENT_OCCURRED_ALIAS), mask);
+	__set_bit(irq, (volatile void *)pda->sn_in_service_ivecs);
+
+	irq_move_irq(data);
+}
+
+struct sn_irq_info *sn_retarget_vector(struct sn_irq_info *sn_irq_info,
+				       nasid_t nasid, int slice)
+{
+	int vector;
+	int cpuid;
+#ifdef CONFIG_SMP
+	int cpuphys;
+#endif
+	int64_t bridge;
+	int local_widget, status;
+	nasid_t local_nasid;
+	struct sn_irq_info *new_irq_info;
+	struct sn_pcibus_provider *pci_provider;
+
+	bridge = (u64) sn_irq_info->irq_bridge;
+	if (!bridge) {
+		return NULL; /* irq is not a device interrupt */
+	}
+
+	local_nasid = NASID_GET(bridge);
+
+	if (local_nasid & 1)
+		local_widget = TIO_SWIN_WIDGETNUM(bridge);
+	else
+		local_widget = SWIN_WIDGETNUM(bridge);
+	vector = sn_irq_info->irq_irq;
+
+	/* Make use of SAL_INTR_REDIRECT if PROM supports it */
+	status = sn_intr_redirect(local_nasid, local_widget, sn_irq_info, nasid, slice);
+	if (!status) {
+		new_irq_info = sn_irq_info;
+		goto finish_up;
+	}
+
+	/*
+	 * PROM does not support SAL_INTR_REDIRECT, or it failed.
+	 * Revert to old method.
+	 */
+	new_irq_info = kmemdup(sn_irq_info, sizeof(struct sn_irq_info),
+			       GFP_ATOMIC);
+	if (new_irq_info == NULL)
+		return NULL;
+
+	/* Free the old PROM new_irq_info structure */
+	sn_intr_free(local_nasid, local_widget, new_irq_info);
+	unregister_intr_pda(new_irq_info);
+
+	/* allocate a new PROM new_irq_info struct */
+	status = sn_intr_alloc(local_nasid, local_widget,
+			       new_irq_info, vector,
+			       nasid, slice);
+
+	/* SAL call failed */
+	if (status) {
+		kfree(new_irq_info);
+		return NULL;
+	}
+
+	register_intr_pda(new_irq_info);
+	spin_lock(&sn_irq_info_lock);
+	list_replace_rcu(&sn_irq_info->list, &new_irq_info->list);
+	spin_unlock(&sn_irq_info_lock);
+	kfree_rcu(sn_irq_info, rcu);
+
+
+finish_up:
+	/* Update kernels new_irq_info with new target info */
+	cpuid = nasid_slice_to_cpuid(new_irq_info->irq_nasid,
+				     new_irq_info->irq_slice);
+	new_irq_info->irq_cpuid = cpuid;
+
+	pci_provider = sn_pci_provider[new_irq_info->irq_bridge_type];
+
+	/*
+	 * If this represents a line interrupt, target it.  If it's
+	 * an msi (irq_int_bit < 0), it's already targeted.
+	 */
+	if (new_irq_info->irq_int_bit >= 0 &&
+	    pci_provider && pci_provider->target_interrupt)
+		(pci_provider->target_interrupt)(new_irq_info);
+
+#ifdef CONFIG_SMP
+	cpuphys = cpu_physical_id(cpuid);
+	set_irq_affinity_info((vector & 0xff), cpuphys, 0);
+#endif
+
+	return new_irq_info;
+}
+
+static int sn_set_affinity_irq(struct irq_data *data,
+			       const struct cpumask *mask, bool force)
+{
+	struct sn_irq_info *sn_irq_info, *sn_irq_info_safe;
+	unsigned int irq = data->irq;
+	nasid_t nasid;
+	int slice;
+
+	nasid = cpuid_to_nasid(cpumask_first_and(mask, cpu_online_mask));
+	slice = cpuid_to_slice(cpumask_first_and(mask, cpu_online_mask));
+
+	list_for_each_entry_safe(sn_irq_info, sn_irq_info_safe,
+				 sn_irq_lh[irq], list)
+		(void)sn_retarget_vector(sn_irq_info, nasid, slice);
+
+	return 0;
+}
+
+#ifdef CONFIG_SMP
+void sn_set_err_irq_affinity(unsigned int irq)
+{
+        /*
+         * On systems which support CPU disabling (SHub2), all error interrupts
+         * are targeted at the boot CPU.
+         */
+        if (is_shub2() && sn_prom_feature_available(PRF_CPU_DISABLE_SUPPORT))
+                set_irq_affinity_info(irq, cpu_physical_id(0), 0);
+}
+#else
+void sn_set_err_irq_affinity(unsigned int irq) { }
+#endif
+
+static void
+sn_mask_irq(struct irq_data *data)
+{
+}
+
+static void
+sn_unmask_irq(struct irq_data *data)
+{
+}
+
+struct irq_chip irq_type_sn = {
+	.name			= "SN hub",
+	.irq_startup		= sn_startup_irq,
+	.irq_shutdown		= sn_shutdown_irq,
+	.irq_enable		= sn_enable_irq,
+	.irq_disable		= sn_disable_irq,
+	.irq_ack		= sn_ack_irq,
+	.irq_mask		= sn_mask_irq,
+	.irq_unmask		= sn_unmask_irq,
+	.irq_set_affinity	= sn_set_affinity_irq
+};
+
+ia64_vector sn_irq_to_vector(int irq)
+{
+	if (irq >= IA64_NUM_VECTORS)
+		return 0;
+	return (ia64_vector)irq;
+}
+
+unsigned int sn_local_vector_to_irq(u8 vector)
+{
+	return (CPU_VECTOR_TO_IRQ(smp_processor_id(), vector));
+}
+
+void sn_irq_init(void)
+{
+	int i;
+
+	ia64_first_device_vector = IA64_SN2_FIRST_DEVICE_VECTOR;
+	ia64_last_device_vector = IA64_SN2_LAST_DEVICE_VECTOR;
+
+	for (i = 0; i < NR_IRQS; i++) {
+		if (irq_get_chip(i) == &no_irq_chip)
+			irq_set_chip(i, &irq_type_sn);
+	}
+}
+
+static void register_intr_pda(struct sn_irq_info *sn_irq_info)
+{
+	int irq = sn_irq_info->irq_irq;
+	int cpu = sn_irq_info->irq_cpuid;
+
+	if (pdacpu(cpu)->sn_last_irq < irq) {
+		pdacpu(cpu)->sn_last_irq = irq;
+	}
+
+	if (pdacpu(cpu)->sn_first_irq == 0 || pdacpu(cpu)->sn_first_irq > irq)
+		pdacpu(cpu)->sn_first_irq = irq;
+}
+
+static void unregister_intr_pda(struct sn_irq_info *sn_irq_info)
+{
+	int irq = sn_irq_info->irq_irq;
+	int cpu = sn_irq_info->irq_cpuid;
+	struct sn_irq_info *tmp_irq_info;
+	int i, foundmatch;
+
+	rcu_read_lock();
+	if (pdacpu(cpu)->sn_last_irq == irq) {
+		foundmatch = 0;
+		for (i = pdacpu(cpu)->sn_last_irq - 1;
+		     i && !foundmatch; i--) {
+			list_for_each_entry_rcu(tmp_irq_info,
+						sn_irq_lh[i],
+						list) {
+				if (tmp_irq_info->irq_cpuid == cpu) {
+					foundmatch = 1;
+					break;
+				}
+			}
+		}
+		pdacpu(cpu)->sn_last_irq = i;
+	}
+
+	if (pdacpu(cpu)->sn_first_irq == irq) {
+		foundmatch = 0;
+		for (i = pdacpu(cpu)->sn_first_irq + 1;
+		     i < NR_IRQS && !foundmatch; i++) {
+			list_for_each_entry_rcu(tmp_irq_info,
+						sn_irq_lh[i],
+						list) {
+				if (tmp_irq_info->irq_cpuid == cpu) {
+					foundmatch = 1;
+					break;
+				}
+			}
+		}
+		pdacpu(cpu)->sn_first_irq = ((i == NR_IRQS) ? 0 : i);
+	}
+	rcu_read_unlock();
+}
+
+void sn_irq_fixup(struct pci_dev *pci_dev, struct sn_irq_info *sn_irq_info)
+{
+	nasid_t nasid = sn_irq_info->irq_nasid;
+	int slice = sn_irq_info->irq_slice;
+	int cpu = nasid_slice_to_cpuid(nasid, slice);
+#ifdef CONFIG_SMP
+	int cpuphys;
+#endif
+
+	pci_dev_get(pci_dev);
+	sn_irq_info->irq_cpuid = cpu;
+	sn_irq_info->irq_pciioinfo = SN_PCIDEV_INFO(pci_dev);
+
+	/* link it into the sn_irq[irq] list */
+	spin_lock(&sn_irq_info_lock);
+	list_add_rcu(&sn_irq_info->list, sn_irq_lh[sn_irq_info->irq_irq]);
+	reserve_irq_vector(sn_irq_info->irq_irq);
+	if (sn_irq_info->irq_int_bit != -1)
+		irq_set_handler(sn_irq_info->irq_irq, handle_level_irq);
+	spin_unlock(&sn_irq_info_lock);
+
+	register_intr_pda(sn_irq_info);
+#ifdef CONFIG_SMP
+	cpuphys = cpu_physical_id(cpu);
+	set_irq_affinity_info(sn_irq_info->irq_irq, cpuphys, 0);
+	/*
+	 * Affinity was set by the PROM, prevent it from
+	 * being reset by the request_irq() path.
+	 */
+	irqd_mark_affinity_was_set(irq_get_irq_data(sn_irq_info->irq_irq));
+#endif
+}
+
+void sn_irq_unfixup(struct pci_dev *pci_dev)
+{
+	struct sn_irq_info *sn_irq_info;
+
+	/* Only cleanup IRQ stuff if this device has a host bus context */
+	if (!SN_PCIDEV_BUSSOFT(pci_dev))
+		return;
+
+	sn_irq_info = SN_PCIDEV_INFO(pci_dev)->pdi_sn_irq_info;
+	if (!sn_irq_info)
+		return;
+	if (!sn_irq_info->irq_irq) {
+		kfree(sn_irq_info);
+		return;
+	}
+
+	unregister_intr_pda(sn_irq_info);
+	spin_lock(&sn_irq_info_lock);
+	list_del_rcu(&sn_irq_info->list);
+	spin_unlock(&sn_irq_info_lock);
+	if (list_empty(sn_irq_lh[sn_irq_info->irq_irq]))
+		free_irq_vector(sn_irq_info->irq_irq);
+	kfree_rcu(sn_irq_info, rcu);
+	pci_dev_put(pci_dev);
+
+}
+
+static inline void
+sn_call_force_intr_provider(struct sn_irq_info *sn_irq_info)
+{
+	struct sn_pcibus_provider *pci_provider;
+
+	pci_provider = sn_pci_provider[sn_irq_info->irq_bridge_type];
+
+	/* Don't force an interrupt if the irq has been disabled */
+	if (!irqd_irq_disabled(irq_get_irq_data(sn_irq_info->irq_irq)) &&
+	    pci_provider && pci_provider->force_interrupt)
+		(*pci_provider->force_interrupt)(sn_irq_info);
+}
+
+/*
+ * Check for lost interrupts.  If the PIC int_status reg. says that
+ * an interrupt has been sent, but not handled, and the interrupt
+ * is not pending in either the cpu irr regs or in the soft irr regs,
+ * and the interrupt is not in service, then the interrupt may have
+ * been lost.  Force an interrupt on that pin.  It is possible that
+ * the interrupt is in flight, so we may generate a spurious interrupt,
+ * but we should never miss a real lost interrupt.
+ */
+static void sn_check_intr(int irq, struct sn_irq_info *sn_irq_info)
+{
+	u64 regval;
+	struct pcidev_info *pcidev_info;
+	struct pcibus_info *pcibus_info;
+
+	/*
+	 * Bridge types attached to TIO (anything but PIC) do not need this WAR
+	 * since they do not target Shub II interrupt registers.  If that
+	 * ever changes, this check needs to accommodate.
+	 */
+	if (sn_irq_info->irq_bridge_type != PCIIO_ASIC_TYPE_PIC)
+		return;
+
+	pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	if (!pcidev_info)
+		return;
+
+	pcibus_info =
+	    (struct pcibus_info *)pcidev_info->pdi_host_pcidev_info->
+	    pdi_pcibus_info;
+	regval = pcireg_intr_status_get(pcibus_info);
+
+	if (!ia64_get_irr(irq_to_vector(irq))) {
+		if (!test_bit(irq, pda->sn_in_service_ivecs)) {
+			regval &= 0xff;
+			if (sn_irq_info->irq_int_bit & regval &
+			    sn_irq_info->irq_last_intr) {
+				regval &= ~(sn_irq_info->irq_int_bit & regval);
+				sn_call_force_intr_provider(sn_irq_info);
+			}
+		}
+	}
+	sn_irq_info->irq_last_intr = regval;
+}
+
+void sn_lb_int_war_check(void)
+{
+	struct sn_irq_info *sn_irq_info;
+	int i;
+
+	if (!sn_ioif_inited || pda->sn_first_irq == 0)
+		return;
+
+	rcu_read_lock();
+	for (i = pda->sn_first_irq; i <= pda->sn_last_irq; i++) {
+		list_for_each_entry_rcu(sn_irq_info, sn_irq_lh[i], list) {
+			sn_check_intr(i, sn_irq_info);
+		}
+	}
+	rcu_read_unlock();
+}
+
+void __init sn_irq_lh_init(void)
+{
+	int i;
+
+	sn_irq_lh = kmalloc(sizeof(struct list_head *) * NR_IRQS, GFP_KERNEL);
+	if (!sn_irq_lh)
+		panic("SN PCI INIT: Failed to allocate memory for PCI init\n");
+
+	for (i = 0; i < NR_IRQS; i++) {
+		sn_irq_lh[i] = kmalloc(sizeof(struct list_head), GFP_KERNEL);
+		if (!sn_irq_lh[i])
+			panic("SN PCI INIT: Failed IRQ memory allocation\n");
+
+		INIT_LIST_HEAD(sn_irq_lh[i]);
+	}
+}
diff --git a/kernel/arch/ia64/sn/kernel/klconflib.c b/kernel/arch/ia64/sn/kernel/klconflib.c
new file mode 100644
index 000000000..87682b48e
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/klconflib.c
@@ -0,0 +1,107 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1992 - 1997, 2000-2004 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/ctype.h>
+#include <linux/string.h>
+#include <linux/kernel.h>
+#include <asm/sn/types.h>
+#include <asm/sn/module.h>
+#include <asm/sn/l1.h>
+
+char brick_types[MAX_BRICK_TYPES + 1] = "cri.xdpn%#=vo^kjbf890123456789...";
+/*
+ * Format a module id for printing.
+ *
+ * There are three possible formats:
+ *
+ *   MODULE_FORMAT_BRIEF	is the brief 6-character format, including
+ *				the actual brick-type as recorded in the 
+ *				moduleid_t, eg. 002c15 for a C-brick, or
+ *				101#17 for a PX-brick.
+ *
+ *   MODULE_FORMAT_LONG		is the hwgraph format, eg. rack/002/bay/15
+ *				of rack/101/bay/17 (note that the brick
+ *				type does not appear in this format).
+ *
+ *   MODULE_FORMAT_LCD		is like MODULE_FORMAT_BRIEF, except that it
+ *				ensures that the module id provided appears
+ *				exactly as it would on the LCD display of
+ *				the corresponding brick, eg. still 002c15
+ *				for a C-brick, but 101p17 for a PX-brick.
+ *
+ * maule (9/13/04):  Removed top-level check for (fmt == MODULE_FORMAT_LCD)
+ * making MODULE_FORMAT_LCD equivalent to MODULE_FORMAT_BRIEF.  It was
+ * decided that all callers should assume the returned string should be what
+ * is displayed on the brick L1 LCD.
+ */
+void
+format_module_id(char *buffer, moduleid_t m, int fmt)
+{
+	int rack, position;
+	unsigned char brickchar;
+
+	rack = MODULE_GET_RACK(m);
+	brickchar = MODULE_GET_BTCHAR(m);
+
+	/* Be sure we use the same brick type character as displayed
+	 * on the brick's LCD
+	 */
+	switch (brickchar) 
+	{
+	case L1_BRICKTYPE_GA:
+	case L1_BRICKTYPE_OPUS_TIO:
+		brickchar = L1_BRICKTYPE_C;
+		break;
+
+	case L1_BRICKTYPE_PX:
+	case L1_BRICKTYPE_PE:
+	case L1_BRICKTYPE_PA:
+	case L1_BRICKTYPE_SA: /* we can move this to the "I's" later
+			       * if that makes more sense
+			       */
+		brickchar = L1_BRICKTYPE_P;
+		break;
+
+	case L1_BRICKTYPE_IX:
+	case L1_BRICKTYPE_IA:
+
+		brickchar = L1_BRICKTYPE_I;
+		break;
+	}
+
+	position = MODULE_GET_BPOS(m);
+
+	if ((fmt == MODULE_FORMAT_BRIEF) || (fmt == MODULE_FORMAT_LCD)) {
+		/* Brief module number format, eg. 002c15 */
+
+		/* Decompress the rack number */
+		*buffer++ = '0' + RACK_GET_CLASS(rack);
+		*buffer++ = '0' + RACK_GET_GROUP(rack);
+		*buffer++ = '0' + RACK_GET_NUM(rack);
+
+		/* Add the brick type */
+		*buffer++ = brickchar;
+	}
+	else if (fmt == MODULE_FORMAT_LONG) {
+		/* Fuller hwgraph format, eg. rack/002/bay/15 */
+
+		strcpy(buffer, "rack" "/");  buffer += strlen(buffer);
+
+		*buffer++ = '0' + RACK_GET_CLASS(rack);
+		*buffer++ = '0' + RACK_GET_GROUP(rack);
+		*buffer++ = '0' + RACK_GET_NUM(rack);
+
+		strcpy(buffer, "/" "bay" "/");  buffer += strlen(buffer);
+	}
+
+	/* Add the bay position, using at least two digits */
+	if (position < 10)
+		*buffer++ = '0';
+	sprintf(buffer, "%d", position);
+}
diff --git a/kernel/arch/ia64/sn/kernel/machvec.c b/kernel/arch/ia64/sn/kernel/machvec.c
new file mode 100644
index 000000000..02bb91558
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/machvec.c
@@ -0,0 +1,11 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2002-2003 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#define MACHVEC_PLATFORM_NAME	sn2
+#define MACHVEC_PLATFORM_HEADER	<asm/machvec_sn2.h>
+#include <asm/machvec_init.h>
diff --git a/kernel/arch/ia64/sn/kernel/mca.c b/kernel/arch/ia64/sn/kernel/mca.c
new file mode 100644
index 000000000..27793f7aa
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/mca.c
@@ -0,0 +1,146 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2000-2006 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/timer.h>
+#include <linux/vmalloc.h>
+#include <linux/mutex.h>
+#include <asm/mca.h>
+#include <asm/sal.h>
+#include <asm/sn/sn_sal.h>
+
+/*
+ * Interval for calling SAL to poll for errors that do NOT cause error
+ * interrupts. SAL will raise a CPEI if any errors are present that
+ * need to be logged.
+ */
+#define CPEI_INTERVAL	(5*HZ)
+
+struct timer_list sn_cpei_timer;
+void sn_init_cpei_timer(void);
+
+/* Printing oemdata from mca uses data that is not passed through SAL, it is
+ * global.  Only one user at a time.
+ */
+static DEFINE_MUTEX(sn_oemdata_mutex);
+static u8 **sn_oemdata;
+static u64 *sn_oemdata_size, sn_oemdata_bufsize;
+
+/*
+ * print_hook
+ *
+ * This function is the callback routine that SAL calls to log error
+ * info for platform errors.  buf is appended to sn_oemdata, resizing as
+ * required.
+ * Note: this is a SAL to OS callback, running under the same rules as the SAL
+ * code.  SAL calls are run with preempt disabled so this routine must not
+ * sleep.  vmalloc can sleep so print_hook cannot resize the output buffer
+ * itself, instead it must set the required size and return to let the caller
+ * resize the buffer then redrive the SAL call.
+ */
+static int print_hook(const char *fmt, ...)
+{
+	char buf[400];
+	int len;
+	va_list args;
+	va_start(args, fmt);
+	vsnprintf(buf, sizeof(buf), fmt, args);
+	va_end(args);
+	len = strlen(buf);
+	if (*sn_oemdata_size + len <= sn_oemdata_bufsize)
+		memcpy(*sn_oemdata + *sn_oemdata_size, buf, len);
+	*sn_oemdata_size += len;
+	return 0;
+}
+
+static void sn_cpei_handler(int irq, void *devid, struct pt_regs *regs)
+{
+	/*
+	 * this function's sole purpose is to call SAL when we receive
+	 * a CE interrupt from SHUB or when the timer routine decides
+	 * we need to call SAL to check for CEs.
+	 */
+
+	/* CALL SAL_LOG_CE */
+
+	ia64_sn_plat_cpei_handler();
+}
+
+static void sn_cpei_timer_handler(unsigned long dummy)
+{
+	sn_cpei_handler(-1, NULL, NULL);
+	mod_timer(&sn_cpei_timer, jiffies + CPEI_INTERVAL);
+}
+
+void sn_init_cpei_timer(void)
+{
+	init_timer(&sn_cpei_timer);
+	sn_cpei_timer.expires = jiffies + CPEI_INTERVAL;
+	sn_cpei_timer.function = sn_cpei_timer_handler;
+	add_timer(&sn_cpei_timer);
+}
+
+static int
+sn_platform_plat_specific_err_print(const u8 * sect_header, u8 ** oemdata,
+				    u64 * oemdata_size)
+{
+	mutex_lock(&sn_oemdata_mutex);
+	sn_oemdata = oemdata;
+	sn_oemdata_size = oemdata_size;
+	sn_oemdata_bufsize = 0;
+	*sn_oemdata_size = PAGE_SIZE;	/* first guess at how much data will be generated */
+	while (*sn_oemdata_size > sn_oemdata_bufsize) {
+		u8 *newbuf = vmalloc(*sn_oemdata_size);
+		if (!newbuf) {
+			mutex_unlock(&sn_oemdata_mutex);
+			printk(KERN_ERR "%s: unable to extend sn_oemdata\n",
+			       __func__);
+			return 1;
+		}
+		vfree(*sn_oemdata);
+		*sn_oemdata = newbuf;
+		sn_oemdata_bufsize = *sn_oemdata_size;
+		*sn_oemdata_size = 0;
+		ia64_sn_plat_specific_err_print(print_hook, (char *)sect_header);
+	}
+	mutex_unlock(&sn_oemdata_mutex);
+	return 0;
+}
+
+/* Callback when userspace salinfo wants to decode oem data via the platform
+ * kernel and/or prom.
+ */
+int sn_salinfo_platform_oemdata(const u8 *sect_header, u8 **oemdata, u64 *oemdata_size)
+{
+	efi_guid_t guid = *(efi_guid_t *)sect_header;
+	int valid = 0;
+	*oemdata_size = 0;
+	vfree(*oemdata);
+	*oemdata = NULL;
+	if (efi_guidcmp(guid, SAL_PLAT_SPECIFIC_ERR_SECT_GUID) == 0) {
+		sal_log_plat_specific_err_info_t *psei = (sal_log_plat_specific_err_info_t *)sect_header;
+		valid = psei->valid.oem_data;
+	} else if (efi_guidcmp(guid, SAL_PLAT_MEM_DEV_ERR_SECT_GUID) == 0) {
+		sal_log_mem_dev_err_info_t *mdei = (sal_log_mem_dev_err_info_t *)sect_header;
+		valid = mdei->valid.oem_data;
+	}
+	if (valid)
+		return sn_platform_plat_specific_err_print(sect_header, oemdata, oemdata_size);
+	else
+		return 0;
+}
+
+static int __init sn_salinfo_init(void)
+{
+	if (ia64_platform_is("sn2"))
+		salinfo_platform_oemdata = &sn_salinfo_platform_oemdata;
+	return 0;
+}
+
+module_init(sn_salinfo_init)
diff --git a/kernel/arch/ia64/sn/kernel/msi_sn.c b/kernel/arch/ia64/sn/kernel/msi_sn.c
new file mode 100644
index 000000000..a0eb27b66
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/msi_sn.c
@@ -0,0 +1,238 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2006 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/irq.h>
+#include <linux/pci.h>
+#include <linux/cpumask.h>
+#include <linux/msi.h>
+#include <linux/slab.h>
+
+#include <asm/sn/addrs.h>
+#include <asm/sn/intr.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/nodepda.h>
+
+struct sn_msi_info {
+	u64 pci_addr;
+	struct sn_irq_info *sn_irq_info;
+};
+
+static struct sn_msi_info sn_msi_info[NR_IRQS];
+
+static struct irq_chip sn_msi_chip;
+
+void sn_teardown_msi_irq(unsigned int irq)
+{
+	nasid_t nasid;
+	int widget;
+	struct pci_dev *pdev;
+	struct pcidev_info *sn_pdev;
+	struct sn_irq_info *sn_irq_info;
+	struct pcibus_bussoft *bussoft;
+	struct sn_pcibus_provider *provider;
+
+	sn_irq_info = sn_msi_info[irq].sn_irq_info;
+	if (sn_irq_info == NULL || sn_irq_info->irq_int_bit >= 0)
+		return;
+
+	sn_pdev = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	pdev = sn_pdev->pdi_linux_pcidev;
+	provider = SN_PCIDEV_BUSPROVIDER(pdev);
+
+	(*provider->dma_unmap)(pdev,
+			       sn_msi_info[irq].pci_addr,
+			       PCI_DMA_FROMDEVICE);
+	sn_msi_info[irq].pci_addr = 0;
+
+	bussoft = SN_PCIDEV_BUSSOFT(pdev);
+	nasid = NASID_GET(bussoft->bs_base);
+	widget = (nasid & 1) ?
+			TIO_SWIN_WIDGETNUM(bussoft->bs_base) :
+			SWIN_WIDGETNUM(bussoft->bs_base);
+
+	sn_intr_free(nasid, widget, sn_irq_info);
+	sn_msi_info[irq].sn_irq_info = NULL;
+
+	destroy_irq(irq);
+}
+
+int sn_setup_msi_irq(struct pci_dev *pdev, struct msi_desc *entry)
+{
+	struct msi_msg msg;
+	int widget;
+	int status;
+	nasid_t nasid;
+	u64 bus_addr;
+	struct sn_irq_info *sn_irq_info;
+	struct pcibus_bussoft *bussoft = SN_PCIDEV_BUSSOFT(pdev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+	int irq;
+
+	if (!entry->msi_attrib.is_64)
+		return -EINVAL;
+
+	if (bussoft == NULL)
+		return -EINVAL;
+
+	if (provider == NULL || provider->dma_map_consistent == NULL)
+		return -EINVAL;
+
+	irq = create_irq();
+	if (irq < 0)
+		return irq;
+
+	/*
+	 * Set up the vector plumbing.  Let the prom (via sn_intr_alloc)
+	 * decide which cpu to direct this msi at by default.
+	 */
+
+	nasid = NASID_GET(bussoft->bs_base);
+	widget = (nasid & 1) ?
+			TIO_SWIN_WIDGETNUM(bussoft->bs_base) :
+			SWIN_WIDGETNUM(bussoft->bs_base);
+
+	sn_irq_info = kzalloc(sizeof(struct sn_irq_info), GFP_KERNEL);
+	if (! sn_irq_info) {
+		destroy_irq(irq);
+		return -ENOMEM;
+	}
+
+	status = sn_intr_alloc(nasid, widget, sn_irq_info, irq, -1, -1);
+	if (status) {
+		kfree(sn_irq_info);
+		destroy_irq(irq);
+		return -ENOMEM;
+	}
+
+	sn_irq_info->irq_int_bit = -1;		/* mark this as an MSI irq */
+	sn_irq_fixup(pdev, sn_irq_info);
+
+	/* Prom probably should fill these in, but doesn't ... */
+	sn_irq_info->irq_bridge_type = bussoft->bs_asic_type;
+	sn_irq_info->irq_bridge = (void *)bussoft->bs_base;
+
+	/*
+	 * Map the xio address into bus space
+	 */
+	bus_addr = (*provider->dma_map_consistent)(pdev,
+					sn_irq_info->irq_xtalkaddr,
+					sizeof(sn_irq_info->irq_xtalkaddr),
+					SN_DMA_MSI|SN_DMA_ADDR_XIO);
+	if (! bus_addr) {
+		sn_intr_free(nasid, widget, sn_irq_info);
+		kfree(sn_irq_info);
+		destroy_irq(irq);
+		return -ENOMEM;
+	}
+
+	sn_msi_info[irq].sn_irq_info = sn_irq_info;
+	sn_msi_info[irq].pci_addr = bus_addr;
+
+	msg.address_hi = (u32)(bus_addr >> 32);
+	msg.address_lo = (u32)(bus_addr & 0x00000000ffffffff);
+
+	/*
+	 * In the SN platform, bit 16 is a "send vector" bit which
+	 * must be present in order to move the vector through the system.
+	 */
+	msg.data = 0x100 + irq;
+
+	irq_set_msi_desc(irq, entry);
+	pci_write_msi_msg(irq, &msg);
+	irq_set_chip_and_handler(irq, &sn_msi_chip, handle_edge_irq);
+
+	return 0;
+}
+
+#ifdef CONFIG_SMP
+static int sn_set_msi_irq_affinity(struct irq_data *data,
+				   const struct cpumask *cpu_mask, bool force)
+{
+	struct msi_msg msg;
+	int slice;
+	nasid_t nasid;
+	u64 bus_addr;
+	struct pci_dev *pdev;
+	struct pcidev_info *sn_pdev;
+	struct sn_irq_info *sn_irq_info;
+	struct sn_irq_info *new_irq_info;
+	struct sn_pcibus_provider *provider;
+	unsigned int cpu, irq = data->irq;
+
+	cpu = cpumask_first_and(cpu_mask, cpu_online_mask);
+	sn_irq_info = sn_msi_info[irq].sn_irq_info;
+	if (sn_irq_info == NULL || sn_irq_info->irq_int_bit >= 0)
+		return -1;
+
+	/*
+	 * Release XIO resources for the old MSI PCI address
+	 */
+
+	__get_cached_msi_msg(data->msi_desc, &msg);
+	sn_pdev = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	pdev = sn_pdev->pdi_linux_pcidev;
+	provider = SN_PCIDEV_BUSPROVIDER(pdev);
+
+	bus_addr = (u64)(msg.address_hi) << 32 | (u64)(msg.address_lo);
+	(*provider->dma_unmap)(pdev, bus_addr, PCI_DMA_FROMDEVICE);
+	sn_msi_info[irq].pci_addr = 0;
+
+	nasid = cpuid_to_nasid(cpu);
+	slice = cpuid_to_slice(cpu);
+
+	new_irq_info = sn_retarget_vector(sn_irq_info, nasid, slice);
+	sn_msi_info[irq].sn_irq_info = new_irq_info;
+	if (new_irq_info == NULL)
+		return -1;
+
+	/*
+	 * Map the xio address into bus space
+	 */
+
+	bus_addr = (*provider->dma_map_consistent)(pdev,
+					new_irq_info->irq_xtalkaddr,
+					sizeof(new_irq_info->irq_xtalkaddr),
+					SN_DMA_MSI|SN_DMA_ADDR_XIO);
+
+	sn_msi_info[irq].pci_addr = bus_addr;
+	msg.address_hi = (u32)(bus_addr >> 32);
+	msg.address_lo = (u32)(bus_addr & 0x00000000ffffffff);
+
+	pci_write_msi_msg(irq, &msg);
+	cpumask_copy(data->affinity, cpu_mask);
+
+	return 0;
+}
+#endif /* CONFIG_SMP */
+
+static void sn_ack_msi_irq(struct irq_data *data)
+{
+	irq_move_irq(data);
+	ia64_eoi();
+}
+
+static int sn_msi_retrigger_irq(struct irq_data *data)
+{
+	unsigned int vector = data->irq;
+	ia64_resend_irq(vector);
+
+	return 1;
+}
+
+static struct irq_chip sn_msi_chip = {
+	.name			= "PCI-MSI",
+	.irq_mask		= pci_msi_mask_irq,
+	.irq_unmask		= pci_msi_unmask_irq,
+	.irq_ack		= sn_ack_msi_irq,
+#ifdef CONFIG_SMP
+	.irq_set_affinity	= sn_set_msi_irq_affinity,
+#endif
+	.irq_retrigger		= sn_msi_retrigger_irq,
+};
diff --git a/kernel/arch/ia64/sn/kernel/pio_phys.S b/kernel/arch/ia64/sn/kernel/pio_phys.S
new file mode 100644
index 000000000..3c7d48d6e
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/pio_phys.S
@@ -0,0 +1,71 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2005 Silicon Graphics, Inc. All rights reserved.
+ *
+ * This file contains macros used to access MMR registers via
+ * uncached physical addresses.
+ *      pio_phys_read_mmr  - read an MMR
+ *      pio_phys_write_mmr - write an MMR
+ *      pio_atomic_phys_write_mmrs - atomically write 1 or 2 MMRs with psr.ic=0
+ *              Second MMR will be skipped if address is NULL
+ *
+ * Addresses passed to these routines should be uncached physical addresses
+ * 	ie., 0x80000....
+ */
+
+
+
+#include <asm/asmmacro.h>
+#include <asm/page.h>
+
+GLOBAL_ENTRY(pio_phys_read_mmr)
+	.prologue
+	.regstk 1,0,0,0
+	.body
+	mov r2=psr
+	rsm psr.i | psr.dt
+	;;
+	srlz.d
+	ld8.acq r8=[r32]
+	;;
+	mov psr.l=r2;;
+	srlz.d
+	br.ret.sptk.many rp
+END(pio_phys_read_mmr)
+
+GLOBAL_ENTRY(pio_phys_write_mmr)
+	.prologue
+	.regstk 2,0,0,0
+	.body
+	mov r2=psr
+	rsm psr.i | psr.dt
+	;;
+	srlz.d
+	st8.rel [r32]=r33
+	;;
+	mov psr.l=r2;;
+	srlz.d
+	br.ret.sptk.many rp
+END(pio_phys_write_mmr)
+
+GLOBAL_ENTRY(pio_atomic_phys_write_mmrs)
+	.prologue
+	.regstk 4,0,0,0
+	.body
+	mov r2=psr
+	cmp.ne p9,p0=r34,r0;
+	rsm psr.i | psr.dt | psr.ic
+	;;
+	srlz.d
+	st8.rel [r32]=r33
+(p9)	st8.rel [r34]=r35
+	;;
+	mov psr.l=r2;;
+	srlz.d
+	br.ret.sptk.many rp
+END(pio_atomic_phys_write_mmrs)
+
+
diff --git a/kernel/arch/ia64/sn/kernel/setup.c b/kernel/arch/ia64/sn/kernel/setup.c
new file mode 100644
index 000000000..5f6b6b48c
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/setup.c
@@ -0,0 +1,775 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1999,2001-2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/delay.h>
+#include <linux/kernel.h>
+#include <linux/kdev_t.h>
+#include <linux/string.h>
+#include <linux/screen_info.h>
+#include <linux/console.h>
+#include <linux/timex.h>
+#include <linux/sched.h>
+#include <linux/ioport.h>
+#include <linux/mm.h>
+#include <linux/serial.h>
+#include <linux/irq.h>
+#include <linux/bootmem.h>
+#include <linux/mmzone.h>
+#include <linux/interrupt.h>
+#include <linux/acpi.h>
+#include <linux/compiler.h>
+#include <linux/root_dev.h>
+#include <linux/nodemask.h>
+#include <linux/pm.h>
+#include <linux/efi.h>
+
+#include <asm/io.h>
+#include <asm/sal.h>
+#include <asm/machvec.h>
+#include <asm/processor.h>
+#include <asm/vga.h>
+#include <asm/setup.h>
+#include <asm/sn/arch.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/pda.h>
+#include <asm/sn/nodepda.h>
+#include <asm/sn/sn_cpuid.h>
+#include <asm/sn/simulator.h>
+#include <asm/sn/leds.h>
+#include <asm/sn/bte.h>
+#include <asm/sn/shub_mmr.h>
+#include <asm/sn/clksupport.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/geo.h>
+#include <asm/sn/sn_feature_sets.h>
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+#include <asm/sn/klconfig.h>
+
+
+DEFINE_PER_CPU(struct pda_s, pda_percpu);
+
+#define MAX_PHYS_MEMORY		(1UL << IA64_MAX_PHYS_BITS)	/* Max physical address supported */
+
+extern void bte_init_node(nodepda_t *, cnodeid_t);
+
+extern void sn_timer_init(void);
+extern unsigned long last_time_offset;
+extern void (*ia64_mark_idle) (int);
+extern void snidle(int);
+
+unsigned long sn_rtc_cycles_per_second;
+EXPORT_SYMBOL(sn_rtc_cycles_per_second);
+
+DEFINE_PER_CPU(struct sn_hub_info_s, __sn_hub_info);
+EXPORT_PER_CPU_SYMBOL(__sn_hub_info);
+
+DEFINE_PER_CPU(short, __sn_cnodeid_to_nasid[MAX_COMPACT_NODES]);
+EXPORT_PER_CPU_SYMBOL(__sn_cnodeid_to_nasid);
+
+DEFINE_PER_CPU(struct nodepda_s *, __sn_nodepda);
+EXPORT_PER_CPU_SYMBOL(__sn_nodepda);
+
+char sn_system_serial_number_string[128];
+EXPORT_SYMBOL(sn_system_serial_number_string);
+u64 sn_partition_serial_number;
+EXPORT_SYMBOL(sn_partition_serial_number);
+u8 sn_partition_id;
+EXPORT_SYMBOL(sn_partition_id);
+u8 sn_system_size;
+EXPORT_SYMBOL(sn_system_size);
+u8 sn_sharing_domain_size;
+EXPORT_SYMBOL(sn_sharing_domain_size);
+u8 sn_coherency_id;
+EXPORT_SYMBOL(sn_coherency_id);
+u8 sn_region_size;
+EXPORT_SYMBOL(sn_region_size);
+int sn_prom_type;	/* 0=hardware, 1=medusa/realprom, 2=medusa/fakeprom */
+
+short physical_node_map[MAX_NUMALINK_NODES];
+static unsigned long sn_prom_features[MAX_PROM_FEATURE_SETS];
+
+EXPORT_SYMBOL(physical_node_map);
+
+int num_cnodes;
+
+static void sn_init_pdas(char **);
+static void build_cnode_tables(void);
+
+static nodepda_t *nodepdaindr[MAX_COMPACT_NODES];
+
+/*
+ * The format of "screen_info" is strange, and due to early i386-setup
+ * code. This is just enough to make the console code think we're on a
+ * VGA color display.
+ */
+struct screen_info sn_screen_info = {
+	.orig_x = 0,
+	.orig_y = 0,
+	.orig_video_mode = 3,
+	.orig_video_cols = 80,
+	.orig_video_ega_bx = 3,
+	.orig_video_lines = 25,
+	.orig_video_isVGA = 1,
+	.orig_video_points = 16
+};
+
+/*
+ * This routine can only be used during init, since
+ * smp_boot_data is an init data structure.
+ * We have to use smp_boot_data.cpu_phys_id to find
+ * the physical id of the processor because the normal
+ * cpu_physical_id() relies on data structures that
+ * may not be initialized yet.
+ */
+
+static int __init pxm_to_nasid(int pxm)
+{
+	int i;
+	int nid;
+
+	nid = pxm_to_node(pxm);
+	for (i = 0; i < num_node_memblks; i++) {
+		if (node_memblk[i].nid == nid) {
+			return NASID_GET(node_memblk[i].start_paddr);
+		}
+	}
+	return -1;
+}
+
+/**
+ * early_sn_setup - early setup routine for SN platforms
+ *
+ * Sets up an initial console to aid debugging.  Intended primarily
+ * for bringup.  See start_kernel() in init/main.c.
+ */
+
+void __init early_sn_setup(void)
+{
+	efi_system_table_t *efi_systab;
+	efi_config_table_t *config_tables;
+	struct ia64_sal_systab *sal_systab;
+	struct ia64_sal_desc_entry_point *ep;
+	char *p;
+	int i, j;
+
+	/*
+	 * Parse enough of the SAL tables to locate the SAL entry point. Since, console
+	 * IO on SN2 is done via SAL calls, early_printk won't work without this.
+	 *
+	 * This code duplicates some of the ACPI table parsing that is in efi.c & sal.c.
+	 * Any changes to those file may have to be made here as well.
+	 */
+	efi_systab = (efi_system_table_t *) __va(ia64_boot_param->efi_systab);
+	config_tables = __va(efi_systab->tables);
+	for (i = 0; i < efi_systab->nr_tables; i++) {
+		if (efi_guidcmp(config_tables[i].guid, SAL_SYSTEM_TABLE_GUID) ==
+		    0) {
+			sal_systab = __va(config_tables[i].table);
+			p = (char *)(sal_systab + 1);
+			for (j = 0; j < sal_systab->entry_count; j++) {
+				if (*p == SAL_DESC_ENTRY_POINT) {
+					ep = (struct ia64_sal_desc_entry_point
+					      *)p;
+					ia64_sal_handler_init(__va
+							      (ep->sal_proc),
+							      __va(ep->gp));
+					return;
+				}
+				p += SAL_DESC_SIZE(*p);
+			}
+		}
+	}
+	/* Uh-oh, SAL not available?? */
+	printk(KERN_ERR "failed to find SAL entry point\n");
+}
+
+extern int platform_intr_list[];
+static int shub_1_1_found;
+
+/*
+ * sn_check_for_wars
+ *
+ * Set flag for enabling shub specific wars
+ */
+
+static inline int is_shub_1_1(int nasid)
+{
+	unsigned long id;
+	int rev;
+
+	if (is_shub2())
+		return 0;
+	id = REMOTE_HUB_L(nasid, SH1_SHUB_ID);
+	rev = (id & SH1_SHUB_ID_REVISION_MASK) >> SH1_SHUB_ID_REVISION_SHFT;
+	return rev <= 2;
+}
+
+static void sn_check_for_wars(void)
+{
+	int cnode;
+
+	if (is_shub2()) {
+		/* none yet */
+	} else {
+		for_each_online_node(cnode) {
+			if (is_shub_1_1(cnodeid_to_nasid(cnode)))
+				shub_1_1_found = 1;
+		}
+	}
+}
+
+/*
+ * Scan the EFI PCDP table (if it exists) for an acceptable VGA console
+ * output device.  If one exists, pick it and set sn_legacy_{io,mem} to
+ * reflect the bus offsets needed to address it.
+ *
+ * Since pcdp support in SN is not supported in the 2.4 kernel (or at least
+ * the one lbs is based on) just declare the needed structs here.
+ *
+ * Reference spec http://www.dig64.org/specifications/DIG64_PCDPv20.pdf
+ *
+ * Returns 0 if no acceptable vga is found, !0 otherwise.
+ *
+ * Note:  This stuff is duped here because Altix requires the PCDP to
+ * locate a usable VGA device due to lack of proper ACPI support.  Structures
+ * could be used from drivers/firmware/pcdp.h, but it was decided that moving
+ * this file to a more public location just for Altix use was undesirable.
+ */
+
+struct hcdp_uart_desc {
+	u8	pad[45];
+};
+
+struct pcdp {
+	u8	signature[4];	/* should be 'HCDP' */
+	u32	length;
+	u8	rev;		/* should be >=3 for pcdp, <3 for hcdp */
+	u8	sum;
+	u8	oem_id[6];
+	u64	oem_tableid;
+	u32	oem_rev;
+	u32	creator_id;
+	u32	creator_rev;
+	u32	num_type0;
+	struct hcdp_uart_desc uart[0];	/* num_type0 of these */
+	/* pcdp descriptors follow */
+}  __attribute__((packed));
+
+struct pcdp_device_desc {
+	u8	type;
+	u8	primary;
+	u16	length;
+	u16	index;
+	/* interconnect specific structure follows */
+	/* device specific structure follows that */
+}  __attribute__((packed));
+
+struct pcdp_interface_pci {
+	u8	type;		/* 1 == pci */
+	u8	reserved;
+	u16	length;
+	u8	segment;
+	u8	bus;
+	u8 	dev;
+	u8	fun;
+	u16	devid;
+	u16	vendid;
+	u32	acpi_interrupt;
+	u64	mmio_tra;
+	u64	ioport_tra;
+	u8	flags;
+	u8	translation;
+}  __attribute__((packed));
+
+struct pcdp_vga_device {
+	u8	num_eas_desc;
+	/* ACPI Extended Address Space Desc follows */
+}  __attribute__((packed));
+
+/* from pcdp_device_desc.primary */
+#define PCDP_PRIMARY_CONSOLE	0x01
+
+/* from pcdp_device_desc.type */
+#define PCDP_CONSOLE_INOUT	0x0
+#define PCDP_CONSOLE_DEBUG	0x1
+#define PCDP_CONSOLE_OUT	0x2
+#define PCDP_CONSOLE_IN		0x3
+#define PCDP_CONSOLE_TYPE_VGA	0x8
+
+#define PCDP_CONSOLE_VGA	(PCDP_CONSOLE_TYPE_VGA | PCDP_CONSOLE_OUT)
+
+/* from pcdp_interface_pci.type */
+#define PCDP_IF_PCI		1
+
+/* from pcdp_interface_pci.translation */
+#define PCDP_PCI_TRANS_IOPORT	0x02
+#define PCDP_PCI_TRANS_MMIO	0x01
+
+#if defined(CONFIG_VT) && defined(CONFIG_VGA_CONSOLE)
+static void
+sn_scan_pcdp(void)
+{
+	u8 *bp;
+	struct pcdp *pcdp;
+	struct pcdp_device_desc device;
+	struct pcdp_interface_pci if_pci;
+	extern struct efi efi;
+
+	if (efi.hcdp == EFI_INVALID_TABLE_ADDR)
+		return;		/* no hcdp/pcdp table */
+
+	pcdp = __va(efi.hcdp);
+
+	if (pcdp->rev < 3)
+		return;		/* only support PCDP (rev >= 3) */
+
+	for (bp = (u8 *)&pcdp->uart[pcdp->num_type0];
+	     bp < (u8 *)pcdp + pcdp->length;
+	     bp += device.length) {
+		memcpy(&device, bp, sizeof(device));
+		if (! (device.primary & PCDP_PRIMARY_CONSOLE))
+			continue;	/* not primary console */
+
+		if (device.type != PCDP_CONSOLE_VGA)
+			continue;	/* not VGA descriptor */
+
+		memcpy(&if_pci, bp+sizeof(device), sizeof(if_pci));
+		if (if_pci.type != PCDP_IF_PCI)
+			continue;	/* not PCI interconnect */
+
+		if (if_pci.translation & PCDP_PCI_TRANS_IOPORT)
+			vga_console_iobase = if_pci.ioport_tra;
+
+		if (if_pci.translation & PCDP_PCI_TRANS_MMIO)
+			vga_console_membase =
+				if_pci.mmio_tra | __IA64_UNCACHED_OFFSET;
+
+		break; /* once we find the primary, we're done */
+	}
+}
+#endif
+
+static unsigned long sn2_rtc_initial;
+
+/**
+ * sn_setup - SN platform setup routine
+ * @cmdline_p: kernel command line
+ *
+ * Handles platform setup for SN machines.  This includes determining
+ * the RTC frequency (via a SAL call), initializing secondary CPUs, and
+ * setting up per-node data areas.  The console is also initialized here.
+ */
+void __init sn_setup(char **cmdline_p)
+{
+	long status, ticks_per_sec, drift;
+	u32 version = sn_sal_rev();
+	extern void sn_cpu_init(void);
+
+	sn2_rtc_initial = rtc_time();
+	ia64_sn_plat_set_error_handling_features();	// obsolete
+	ia64_sn_set_os_feature(OSF_MCA_SLV_TO_OS_INIT_SLV);
+	ia64_sn_set_os_feature(OSF_FEAT_LOG_SBES);
+	/*
+	 * Note: The calls to notify the PROM of ACPI and PCI Segment
+	 *	 support must be done prior to acpi_load_tables(), as
+	 *	 an ACPI capable PROM will rebuild the DSDT as result
+	 *	 of the call.
+	 */
+	ia64_sn_set_os_feature(OSF_PCISEGMENT_ENABLE);
+	ia64_sn_set_os_feature(OSF_ACPI_ENABLE);
+
+	/* Load the new DSDT and SSDT tables into the global table list. */
+	acpi_table_init();
+
+#if defined(CONFIG_VT) && defined(CONFIG_VGA_CONSOLE)
+	/*
+	 * Handle SN vga console.
+	 *
+	 * SN systems do not have enough ACPI table information
+	 * being passed from prom to identify VGA adapters and the legacy
+	 * addresses to access them.  Until that is done, SN systems rely
+	 * on the PCDP table to identify the primary VGA console if one
+	 * exists.
+	 *
+	 * However, kernel PCDP support is optional, and even if it is built
+	 * into the kernel, it will not be used if the boot cmdline contains
+	 * console= directives.
+	 *
+	 * So, to work around this mess, we duplicate some of the PCDP code
+	 * here so that the primary VGA console (as defined by PCDP) will
+	 * work on SN systems even if a different console (e.g. serial) is
+	 * selected on the boot line (or CONFIG_EFI_PCDP is off).
+	 */
+
+	if (! vga_console_membase)
+		sn_scan_pcdp();
+
+	/*
+	 *	Setup legacy IO space.
+	 *	vga_console_iobase maps to PCI IO Space address 0 on the
+	 * 	bus containing the VGA console.
+	 */
+	if (vga_console_iobase) {
+		io_space[0].mmio_base =
+			(unsigned long) ioremap(vga_console_iobase, 0);
+		io_space[0].sparse = 0;
+	}
+
+	if (vga_console_membase) {
+		/* usable vga ... make tty0 the preferred default console */
+		if (!strstr(*cmdline_p, "console="))
+			add_preferred_console("tty", 0, NULL);
+	} else {
+		printk(KERN_DEBUG "SGI: Disabling VGA console\n");
+		if (!strstr(*cmdline_p, "console="))
+			add_preferred_console("ttySG", 0, NULL);
+#ifdef CONFIG_DUMMY_CONSOLE
+		conswitchp = &dummy_con;
+#else
+		conswitchp = NULL;
+#endif				/* CONFIG_DUMMY_CONSOLE */
+	}
+#endif				/* def(CONFIG_VT) && def(CONFIG_VGA_CONSOLE) */
+
+	MAX_DMA_ADDRESS = PAGE_OFFSET + MAX_PHYS_MEMORY;
+
+	/*
+	 * Build the tables for managing cnodes.
+	 */
+	build_cnode_tables();
+
+	status =
+	    ia64_sal_freq_base(SAL_FREQ_BASE_REALTIME_CLOCK, &ticks_per_sec,
+			       &drift);
+	if (status != 0 || ticks_per_sec < 100000) {
+		printk(KERN_WARNING
+		       "unable to determine platform RTC clock frequency, guessing.\n");
+		/* PROM gives wrong value for clock freq. so guess */
+		sn_rtc_cycles_per_second = 1000000000000UL / 30000UL;
+	} else
+		sn_rtc_cycles_per_second = ticks_per_sec;
+
+	platform_intr_list[ACPI_INTERRUPT_CPEI] = IA64_CPE_VECTOR;
+
+	printk("SGI SAL version %x.%02x\n", version >> 8, version & 0x00FF);
+
+	/*
+	 * we set the default root device to /dev/hda
+	 * to make simulation easy
+	 */
+	ROOT_DEV = Root_HDA1;
+
+	/*
+	 * Create the PDAs and NODEPDAs for all the cpus.
+	 */
+	sn_init_pdas(cmdline_p);
+
+	ia64_mark_idle = &snidle;
+
+	/*
+	 * For the bootcpu, we do this here. All other cpus will make the
+	 * call as part of cpu_init in slave cpu initialization.
+	 */
+	sn_cpu_init();
+
+#ifdef CONFIG_SMP
+	init_smp_config();
+#endif
+	screen_info = sn_screen_info;
+
+	sn_timer_init();
+
+	/*
+	 * set pm_power_off to a SAL call to allow
+	 * sn machines to power off. The SAL call can be replaced
+	 * by an ACPI interface call when ACPI is fully implemented
+	 * for sn.
+	 */
+	pm_power_off = ia64_sn_power_down;
+	current->thread.flags |= IA64_THREAD_MIGRATION;
+}
+
+/**
+ * sn_init_pdas - setup node data areas
+ *
+ * One time setup for Node Data Area.  Called by sn_setup().
+ */
+static void __init sn_init_pdas(char **cmdline_p)
+{
+	cnodeid_t cnode;
+
+	/*
+	 * Allocate & initialize the nodepda for each node.
+	 */
+	for_each_online_node(cnode) {
+		nodepdaindr[cnode] =
+		    alloc_bootmem_node(NODE_DATA(cnode), sizeof(nodepda_t));
+		memset(nodepdaindr[cnode]->phys_cpuid, -1,
+		    sizeof(nodepdaindr[cnode]->phys_cpuid));
+		spin_lock_init(&nodepdaindr[cnode]->ptc_lock);
+	}
+
+	/*
+	 * Allocate & initialize nodepda for TIOs.  For now, put them on node 0.
+	 */
+	for (cnode = num_online_nodes(); cnode < num_cnodes; cnode++)
+		nodepdaindr[cnode] =
+		    alloc_bootmem_node(NODE_DATA(0), sizeof(nodepda_t));
+
+	/*
+	 * Now copy the array of nodepda pointers to each nodepda.
+	 */
+	for (cnode = 0; cnode < num_cnodes; cnode++)
+		memcpy(nodepdaindr[cnode]->pernode_pdaindr, nodepdaindr,
+		       sizeof(nodepdaindr));
+
+	/*
+	 * Set up IO related platform-dependent nodepda fields.
+	 * The following routine actually sets up the hubinfo struct
+	 * in nodepda.
+	 */
+	for_each_online_node(cnode) {
+		bte_init_node(nodepdaindr[cnode], cnode);
+	}
+
+	/*
+	 * Initialize the per node hubdev.  This includes IO Nodes and
+	 * headless/memless nodes.
+	 */
+	for (cnode = 0; cnode < num_cnodes; cnode++) {
+		hubdev_init_node(nodepdaindr[cnode], cnode);
+	}
+}
+
+/**
+ * sn_cpu_init - initialize per-cpu data areas
+ * @cpuid: cpuid of the caller
+ *
+ * Called during cpu initialization on each cpu as it starts.
+ * Currently, initializes the per-cpu data area for SNIA.
+ * Also sets up a few fields in the nodepda.  Also known as
+ * platform_cpu_init() by the ia64 machvec code.
+ */
+void sn_cpu_init(void)
+{
+	int cpuid;
+	int cpuphyid;
+	int nasid;
+	int subnode;
+	int slice;
+	int cnode;
+	int i;
+	static int wars_have_been_checked, set_cpu0_number;
+
+	cpuid = smp_processor_id();
+	if (cpuid == 0 && IS_MEDUSA()) {
+		if (ia64_sn_is_fake_prom())
+			sn_prom_type = 2;
+		else
+			sn_prom_type = 1;
+		printk(KERN_INFO "Running on medusa with %s PROM\n",
+		       (sn_prom_type == 1) ? "real" : "fake");
+	}
+
+	memset(pda, 0, sizeof(*pda));
+	if (ia64_sn_get_sn_info(0, &sn_hub_info->shub2,
+				&sn_hub_info->nasid_bitmask,
+				&sn_hub_info->nasid_shift,
+				&sn_system_size, &sn_sharing_domain_size,
+				&sn_partition_id, &sn_coherency_id,
+				&sn_region_size))
+		BUG();
+	sn_hub_info->as_shift = sn_hub_info->nasid_shift - 2;
+
+	/*
+	 * Don't check status. The SAL call is not supported on all PROMs
+	 * but a failure is harmless.
+	 * Architecturally, cpu_init is always called twice on cpu 0. We
+	 * should set cpu_number on cpu 0 once.
+	 */
+	if (cpuid == 0) {
+		if (!set_cpu0_number) {
+			(void) ia64_sn_set_cpu_number(cpuid);
+			set_cpu0_number = 1;
+		}
+	} else
+		(void) ia64_sn_set_cpu_number(cpuid);
+
+	/*
+	 * The boot cpu makes this call again after platform initialization is
+	 * complete.
+	 */
+	if (nodepdaindr[0] == NULL)
+		return;
+
+	for (i = 0; i < MAX_PROM_FEATURE_SETS; i++)
+		if (ia64_sn_get_prom_feature_set(i, &sn_prom_features[i]) != 0)
+			break;
+
+	cpuphyid = get_sapicid();
+
+	if (ia64_sn_get_sapic_info(cpuphyid, &nasid, &subnode, &slice))
+		BUG();
+
+	for (i=0; i < MAX_NUMNODES; i++) {
+		if (nodepdaindr[i]) {
+			nodepdaindr[i]->phys_cpuid[cpuid].nasid = nasid;
+			nodepdaindr[i]->phys_cpuid[cpuid].slice = slice;
+			nodepdaindr[i]->phys_cpuid[cpuid].subnode = subnode;
+		}
+	}
+
+	cnode = nasid_to_cnodeid(nasid);
+
+	__this_cpu_write(__sn_nodepda, nodepdaindr[cnode]);
+
+	pda->led_address =
+	    (typeof(pda->led_address)) (LED0 + (slice << LED_CPU_SHIFT));
+	pda->led_state = LED_ALWAYS_SET;
+	pda->hb_count = HZ / 2;
+	pda->hb_state = 0;
+	pda->idle_flag = 0;
+
+	if (cpuid != 0) {
+		/* copy cpu 0's sn_cnodeid_to_nasid table to this cpu's */
+		memcpy(sn_cnodeid_to_nasid,
+		       (&per_cpu(__sn_cnodeid_to_nasid, 0)),
+		       sizeof(__ia64_per_cpu_var(__sn_cnodeid_to_nasid)));
+	}
+
+	/*
+	 * Check for WARs.
+	 * Only needs to be done once, on BSP.
+	 * Has to be done after loop above, because it uses this cpu's
+	 * sn_cnodeid_to_nasid table which was just initialized if this
+	 * isn't cpu 0.
+	 * Has to be done before assignment below.
+	 */
+	if (!wars_have_been_checked) {
+		sn_check_for_wars();
+		wars_have_been_checked = 1;
+	}
+	sn_hub_info->shub_1_1_found = shub_1_1_found;
+
+	/*
+	 * Set up addresses of PIO/MEM write status registers.
+	 */
+	{
+		u64 pio1[] = {SH1_PIO_WRITE_STATUS_0, 0, SH1_PIO_WRITE_STATUS_1, 0};
+		u64 pio2[] = {SH2_PIO_WRITE_STATUS_0, SH2_PIO_WRITE_STATUS_2,
+			SH2_PIO_WRITE_STATUS_1, SH2_PIO_WRITE_STATUS_3};
+		u64 *pio;
+		pio = is_shub1() ? pio1 : pio2;
+		pda->pio_write_status_addr =
+		   (volatile unsigned long *)GLOBAL_MMR_ADDR(nasid, pio[slice]);
+		pda->pio_write_status_val = is_shub1() ? SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK : 0;
+	}
+
+	/*
+	 * WAR addresses for SHUB 1.x.
+	 */
+	if (local_node_data->active_cpu_count++ == 0 && is_shub1()) {
+		int buddy_nasid;
+		buddy_nasid =
+		    cnodeid_to_nasid(numa_node_id() ==
+				     num_online_nodes() - 1 ? 0 : numa_node_id() + 1);
+		pda->pio_shub_war_cam_addr =
+		    (volatile unsigned long *)GLOBAL_MMR_ADDR(nasid,
+							      SH1_PI_CAM_CONTROL);
+	}
+}
+
+/*
+ * Build tables for converting between NASIDs and cnodes.
+ */
+static inline int __init board_needs_cnode(int type)
+{
+	return (type == KLTYPE_SNIA || type == KLTYPE_TIO);
+}
+
+void __init build_cnode_tables(void)
+{
+	int nasid;
+	int node;
+	lboard_t *brd;
+
+	memset(physical_node_map, -1, sizeof(physical_node_map));
+	memset(sn_cnodeid_to_nasid, -1,
+			sizeof(__ia64_per_cpu_var(__sn_cnodeid_to_nasid)));
+
+	/*
+	 * First populate the tables with C/M bricks. This ensures that
+	 * cnode == node for all C & M bricks.
+	 */
+	for_each_online_node(node) {
+		nasid = pxm_to_nasid(node_to_pxm(node));
+		sn_cnodeid_to_nasid[node] = nasid;
+		physical_node_map[nasid] = node;
+	}
+
+	/*
+	 * num_cnodes is total number of C/M/TIO bricks. Because of the 256 node
+	 * limit on the number of nodes, we can't use the generic node numbers 
+	 * for this. Note that num_cnodes is incremented below as TIOs or
+	 * headless/memoryless nodes are discovered.
+	 */
+	num_cnodes = num_online_nodes();
+
+	/* fakeprom does not support klgraph */
+	if (IS_RUNNING_ON_FAKE_PROM())
+		return;
+
+	/* Find TIOs & headless/memoryless nodes and add them to the tables */
+	for_each_online_node(node) {
+		kl_config_hdr_t *klgraph_header;
+		nasid = cnodeid_to_nasid(node);
+		klgraph_header = ia64_sn_get_klconfig_addr(nasid);
+		BUG_ON(klgraph_header == NULL);
+		brd = NODE_OFFSET_TO_LBOARD(nasid, klgraph_header->ch_board_info);
+		while (brd) {
+			if (board_needs_cnode(brd->brd_type) && physical_node_map[brd->brd_nasid] < 0) {
+				sn_cnodeid_to_nasid[num_cnodes] = brd->brd_nasid;
+				physical_node_map[brd->brd_nasid] = num_cnodes++;
+			}
+			brd = find_lboard_next(brd);
+		}
+	}
+}
+
+int
+nasid_slice_to_cpuid(int nasid, int slice)
+{
+	long cpu;
+
+	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
+		if (cpuid_to_nasid(cpu) == nasid &&
+					cpuid_to_slice(cpu) == slice)
+			return cpu;
+
+	return -1;
+}
+
+int sn_prom_feature_available(int id)
+{
+	if (id >= BITS_PER_LONG * MAX_PROM_FEATURE_SETS)
+		return 0;
+	return test_bit(id, sn_prom_features);
+}
+
+void
+sn_kernel_launch_event(void)
+{
+	/* ignore status until we understand possible failure, if any*/
+	if (ia64_sn_kernel_launch_event())
+		printk(KERN_ERR "KEXEC is not supported in this PROM, Please update the PROM.\n");
+}
+EXPORT_SYMBOL(sn_prom_feature_available);
+
diff --git a/kernel/arch/ia64/sn/kernel/sn2/Makefile b/kernel/arch/ia64/sn/kernel/sn2/Makefile
new file mode 100644
index 000000000..3d09108d4
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/Makefile
@@ -0,0 +1,15 @@
+# arch/ia64/sn/kernel/sn2/Makefile
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# Copyright (C) 1999,2001-2002 Silicon Graphics, Inc. All rights reserved.
+#
+# sn2 specific kernel files
+#
+
+ccflags-y := -Iarch/ia64/sn/include
+
+obj-y += cache.o io.o ptc_deadlock.o sn2_smp.o sn_proc_fs.o \
+	 prominfo_proc.o timer.o timer_interrupt.o sn_hwperf.o
diff --git a/kernel/arch/ia64/sn/kernel/sn2/cache.c b/kernel/arch/ia64/sn/kernel/sn2/cache.c
new file mode 100644
index 000000000..2862cb330
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/cache.c
@@ -0,0 +1,41 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ * 
+ * Copyright (C) 2001-2003, 2006 Silicon Graphics, Inc. All rights reserved.
+ *
+ */
+#include <linux/module.h>
+#include <asm/pgalloc.h>
+#include <asm/sn/arch.h>
+
+/**
+ * sn_flush_all_caches - flush a range of address from all caches (incl. L4)
+ * @flush_addr: identity mapped region 7 address to start flushing
+ * @bytes: number of bytes to flush
+ *
+ * Flush a range of addresses from all caches including L4. 
+ * All addresses fully or partially contained within 
+ * @flush_addr to @flush_addr + @bytes are flushed
+ * from all caches.
+ */
+void
+sn_flush_all_caches(long flush_addr, long bytes)
+{
+	unsigned long addr = flush_addr;
+
+	/* SHub1 requires a cached address */
+	if (is_shub1() && (addr & RGN_BITS) == RGN_BASE(RGN_UNCACHED))
+		addr = (addr - RGN_BASE(RGN_UNCACHED)) + RGN_BASE(RGN_KERNEL);
+
+	flush_icache_range(addr, addr + bytes);
+	/*
+	 * The last call may have returned before the caches
+	 * were actually flushed, so we call it again to make
+	 * sure.
+	 */
+	flush_icache_range(addr, addr + bytes);
+	mb();
+}
+EXPORT_SYMBOL(sn_flush_all_caches);
diff --git a/kernel/arch/ia64/sn/kernel/sn2/io.c b/kernel/arch/ia64/sn/kernel/sn2/io.c
new file mode 100644
index 000000000..a12c0586d
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/io.c
@@ -0,0 +1,101 @@
+/* 
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2003 Silicon Graphics, Inc. All rights reserved.
+ *
+ * The generic kernel requires function pointers to these routines, so
+ * we wrap the inlines from asm/ia64/sn/sn2/io.h here.
+ */
+
+#include <asm/sn/io.h>
+
+#ifdef CONFIG_IA64_GENERIC
+
+#undef __sn_inb
+#undef __sn_inw
+#undef __sn_inl
+#undef __sn_outb
+#undef __sn_outw
+#undef __sn_outl
+#undef __sn_readb
+#undef __sn_readw
+#undef __sn_readl
+#undef __sn_readq
+#undef __sn_readb_relaxed
+#undef __sn_readw_relaxed
+#undef __sn_readl_relaxed
+#undef __sn_readq_relaxed
+
+unsigned int __sn_inb(unsigned long port)
+{
+	return ___sn_inb(port);
+}
+
+unsigned int __sn_inw(unsigned long port)
+{
+	return ___sn_inw(port);
+}
+
+unsigned int __sn_inl(unsigned long port)
+{
+	return ___sn_inl(port);
+}
+
+void __sn_outb(unsigned char val, unsigned long port)
+{
+	___sn_outb(val, port);
+}
+
+void __sn_outw(unsigned short val, unsigned long port)
+{
+	___sn_outw(val, port);
+}
+
+void __sn_outl(unsigned int val, unsigned long port)
+{
+	___sn_outl(val, port);
+}
+
+unsigned char __sn_readb(void __iomem *addr)
+{
+	return ___sn_readb(addr);
+}
+
+unsigned short __sn_readw(void __iomem *addr)
+{
+	return ___sn_readw(addr);
+}
+
+unsigned int __sn_readl(void __iomem *addr)
+{
+	return ___sn_readl(addr);
+}
+
+unsigned long __sn_readq(void __iomem *addr)
+{
+	return ___sn_readq(addr);
+}
+
+unsigned char __sn_readb_relaxed(void __iomem *addr)
+{
+	return ___sn_readb_relaxed(addr);
+}
+
+unsigned short __sn_readw_relaxed(void __iomem *addr)
+{
+	return ___sn_readw_relaxed(addr);
+}
+
+unsigned int __sn_readl_relaxed(void __iomem *addr)
+{
+	return ___sn_readl_relaxed(addr);
+}
+
+unsigned long __sn_readq_relaxed(void __iomem *addr)
+{
+	return ___sn_readq_relaxed(addr);
+}
+
+#endif
diff --git a/kernel/arch/ia64/sn/kernel/sn2/prominfo_proc.c b/kernel/arch/ia64/sn/kernel/sn2/prominfo_proc.c
new file mode 100644
index 000000000..ec4de2b09
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/prominfo_proc.c
@@ -0,0 +1,231 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1999,2001-2004, 2006 Silicon Graphics, Inc.  All Rights Reserved.
+ *
+ * Module to export the system's Firmware Interface Tables, including
+ * PROM revision numbers and banners, in /proc
+ */
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <linux/nodemask.h>
+#include <asm/io.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/sn_cpuid.h>
+#include <asm/sn/addrs.h>
+
+MODULE_DESCRIPTION("PROM version reporting for /proc");
+MODULE_AUTHOR("Chad Talbott");
+MODULE_LICENSE("GPL");
+
+/* Standard Intel FIT entry types */
+#define FIT_ENTRY_FIT_HEADER	0x00	/* FIT header entry */
+#define FIT_ENTRY_PAL_B		0x01	/* PAL_B entry */
+/* Entries 0x02 through 0x0D reserved by Intel */
+#define FIT_ENTRY_PAL_A_PROC	0x0E	/* Processor-specific PAL_A entry */
+#define FIT_ENTRY_PAL_A		0x0F	/* PAL_A entry, same as... */
+#define FIT_ENTRY_PAL_A_GEN	0x0F	/* ...Generic PAL_A entry */
+#define FIT_ENTRY_UNUSED	0x7F	/* Unused (reserved by Intel?) */
+/* OEM-defined entries range from 0x10 to 0x7E. */
+#define FIT_ENTRY_SAL_A		0x10	/* SAL_A entry */
+#define FIT_ENTRY_SAL_B		0x11	/* SAL_B entry */
+#define FIT_ENTRY_SALRUNTIME	0x12	/* SAL runtime entry */
+#define FIT_ENTRY_EFI		0x1F	/* EFI entry */
+#define FIT_ENTRY_FPSWA		0x20	/* embedded fpswa entry */
+#define FIT_ENTRY_VMLINUX	0x21	/* embedded vmlinux entry */
+
+#define FIT_MAJOR_SHIFT	(32 + 8)
+#define FIT_MAJOR_MASK	((1 << 8) - 1)
+#define FIT_MINOR_SHIFT	32
+#define FIT_MINOR_MASK	((1 << 8) - 1)
+
+#define FIT_MAJOR(q)	\
+	((unsigned) ((q) >> FIT_MAJOR_SHIFT) & FIT_MAJOR_MASK)
+#define FIT_MINOR(q)	\
+	((unsigned) ((q) >> FIT_MINOR_SHIFT) & FIT_MINOR_MASK)
+
+#define FIT_TYPE_SHIFT	(32 + 16)
+#define FIT_TYPE_MASK	((1 << 7) - 1)
+
+#define FIT_TYPE(q)	\
+	((unsigned) ((q) >> FIT_TYPE_SHIFT) & FIT_TYPE_MASK)
+
+struct fit_type_map_t {
+	unsigned char type;
+	const char *name;
+};
+
+static const struct fit_type_map_t fit_entry_types[] = {
+	{FIT_ENTRY_FIT_HEADER, "FIT Header"},
+	{FIT_ENTRY_PAL_A_GEN, "Generic PAL_A"},
+	{FIT_ENTRY_PAL_A_PROC, "Processor-specific PAL_A"},
+	{FIT_ENTRY_PAL_A, "PAL_A"},
+	{FIT_ENTRY_PAL_B, "PAL_B"},
+	{FIT_ENTRY_SAL_A, "SAL_A"},
+	{FIT_ENTRY_SAL_B, "SAL_B"},
+	{FIT_ENTRY_SALRUNTIME, "SAL runtime"},
+	{FIT_ENTRY_EFI, "EFI"},
+	{FIT_ENTRY_VMLINUX, "Embedded Linux"},
+	{FIT_ENTRY_FPSWA, "Embedded FPSWA"},
+	{FIT_ENTRY_UNUSED, "Unused"},
+	{0xff, "Error"},
+};
+
+static const char *fit_type_name(unsigned char type)
+{
+	struct fit_type_map_t const *mapp;
+
+	for (mapp = fit_entry_types; mapp->type != 0xff; mapp++)
+		if (type == mapp->type)
+			return mapp->name;
+
+	if ((type > FIT_ENTRY_PAL_A) && (type < FIT_ENTRY_UNUSED))
+		return "OEM type";
+	if ((type > FIT_ENTRY_PAL_B) && (type < FIT_ENTRY_PAL_A))
+		return "Reserved";
+
+	return "Unknown type";
+}
+
+static int
+get_fit_entry(unsigned long nasid, int index, unsigned long *fentry,
+	      char *banner, int banlen)
+{
+	return ia64_sn_get_fit_compt(nasid, index, fentry, banner, banlen);
+}
+
+
+/*
+ * These two routines display the FIT table for each node.
+ */
+static void dump_fit_entry(struct seq_file *m, unsigned long *fentry)
+{
+	unsigned type;
+
+	type = FIT_TYPE(fentry[1]);
+	seq_printf(m, "%02x %-25s %x.%02x %016lx %u\n",
+		   type,
+		   fit_type_name(type),
+		   FIT_MAJOR(fentry[1]), FIT_MINOR(fentry[1]),
+		   fentry[0],
+		   /* mult by sixteen to get size in bytes */
+		   (unsigned)(fentry[1] & 0xffffff) * 16);
+}
+
+
+/*
+ * We assume that the fit table will be small enough that we can print
+ * the whole thing into one page.  (This is true for our default 16kB
+ * pages -- each entry is about 60 chars wide when printed.)  I read
+ * somewhere that the maximum size of the FIT is 128 entries, so we're
+ * OK except for 4kB pages (and no one is going to do that on SN
+ * anyway).
+ */
+static int proc_fit_show(struct seq_file *m, void *v)
+{
+	unsigned long nasid = (unsigned long)m->private;
+	unsigned long fentry[2];
+	int index;
+
+	for (index=0;;index++) {
+		BUG_ON(index * 60 > PAGE_SIZE);
+		if (get_fit_entry(nasid, index, fentry, NULL, 0))
+			break;
+		dump_fit_entry(m, fentry);
+	}
+	return 0;
+}
+
+static int proc_fit_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, proc_fit_show, PDE_DATA(inode));
+}
+
+static const struct file_operations proc_fit_fops = {
+	.open		= proc_fit_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static int proc_version_show(struct seq_file *m, void *v)
+{
+	unsigned long nasid = (unsigned long)m->private;
+	unsigned long fentry[2];
+	char banner[128];
+	int index;
+
+	for (index = 0; ; index++) {
+		if (get_fit_entry(nasid, index, fentry, banner,
+				  sizeof(banner)))
+			return 0;
+		if (FIT_TYPE(fentry[1]) == FIT_ENTRY_SAL_A)
+			break;
+	}
+
+	seq_printf(m, "%x.%02x\n", FIT_MAJOR(fentry[1]), FIT_MINOR(fentry[1]));
+
+	if (banner[0])
+		seq_printf(m, "%s\n", banner);
+	return 0;
+}
+
+static int proc_version_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, proc_version_show, PDE_DATA(inode));
+}
+
+static const struct file_operations proc_version_fops = {
+	.open		= proc_version_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+/* module entry points */
+int __init prominfo_init(void);
+void __exit prominfo_exit(void);
+
+module_init(prominfo_init);
+module_exit(prominfo_exit);
+
+#define NODE_NAME_LEN 11
+
+int __init prominfo_init(void)
+{
+	struct proc_dir_entry *sgi_prominfo_entry;
+	cnodeid_t cnodeid;
+
+	if (!ia64_platform_is("sn2"))
+		return 0;
+
+	sgi_prominfo_entry = proc_mkdir("sgi_prominfo", NULL);
+	if (!sgi_prominfo_entry)
+		return -ENOMEM;
+
+	for_each_online_node(cnodeid) {
+		struct proc_dir_entry *dir;
+		unsigned long nasid;
+		char name[NODE_NAME_LEN];
+
+		sprintf(name, "node%d", cnodeid);
+		dir = proc_mkdir(name, sgi_prominfo_entry);
+		if (!dir)
+			continue;
+		nasid = cnodeid_to_nasid(cnodeid);
+		proc_create_data("fit", 0, dir,
+				 &proc_fit_fops, (void *)nasid);
+		proc_create_data("version", 0, dir,
+				 &proc_version_fops, (void *)nasid);
+	}
+	return 0;
+}
+
+void __exit prominfo_exit(void)
+{
+	remove_proc_subtree("sgi_prominfo", NULL);
+}
diff --git a/kernel/arch/ia64/sn/kernel/sn2/ptc_deadlock.S b/kernel/arch/ia64/sn/kernel/sn2/ptc_deadlock.S
new file mode 100644
index 000000000..bebbcc4f8
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/ptc_deadlock.S
@@ -0,0 +1,92 @@
+/* 
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <asm/types.h>
+#include <asm/sn/shub_mmr.h>
+
+#define DEADLOCKBIT	SH_PIO_WRITE_STATUS_WRITE_DEADLOCK_SHFT
+#define WRITECOUNTMASK	SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK
+#define ALIAS_OFFSET	8
+
+
+	.global	sn2_ptc_deadlock_recovery_core
+	.proc  	sn2_ptc_deadlock_recovery_core
+
+sn2_ptc_deadlock_recovery_core:
+	.regstk 6,0,0,0
+
+	ptc0  	 = in0
+	data0 	 = in1
+	ptc1  	 = in2
+	data1 	 = in3
+	piowc 	 = in4
+	zeroval  = in5
+	piowcphy = r30
+	psrsave  = r2
+	scr1	 = r16
+	scr2	 = r17
+	mask	 = r18
+
+
+	extr.u	piowcphy=piowc,0,61;;	// Convert piowc to uncached physical address
+	dep	piowcphy=-1,piowcphy,63,1
+	movl	mask=WRITECOUNTMASK
+	mov	r8=r0
+
+1:
+	cmp.ne  p8,p9=r0,ptc1		// Test for shub type (ptc1 non-null on shub1)
+					// p8 = 1 if shub1, p9 = 1 if shub2
+
+	add	scr2=ALIAS_OFFSET,piowc	// Address of WRITE_STATUS alias register 
+	mov	scr1=7;;		// Clear DEADLOCK, WRITE_ERROR, MULTI_WRITE_ERROR
+(p8)	st8.rel	[scr2]=scr1;;
+(p9)	ld8.acq	scr1=[scr2];;
+
+5:	ld8.acq	scr1=[piowc];;		// Wait for PIOs to complete.
+	hint	@pause
+	and	scr2=scr1,mask;;	// mask of writecount bits
+	cmp.ne	p6,p0=zeroval,scr2
+(p6)	br.cond.sptk 5b
+	
+
+
+	////////////// BEGIN PHYSICAL MODE ////////////////////
+	mov psrsave=psr			// Disable IC (no PMIs)
+	rsm psr.i | psr.dt | psr.ic;;
+	srlz.i;;
+
+	st8.rel [ptc0]=data0		// Write PTC0 & wait for completion.
+
+5:	ld8.acq	scr1=[piowcphy];;	// Wait for PIOs to complete.
+	hint	@pause
+	and	scr2=scr1,mask;;	// mask of writecount bits
+	cmp.ne	p6,p0=zeroval,scr2
+(p6)	br.cond.sptk 5b;;
+
+	tbit.nz	p8,p7=scr1,DEADLOCKBIT;;// Test for DEADLOCK
+(p7)	cmp.ne p7,p0=r0,ptc1;;		// Test for non-null ptc1
+	
+(p7)	st8.rel [ptc1]=data1;;		// Now write PTC1.
+
+5:	ld8.acq	scr1=[piowcphy];;	// Wait for PIOs to complete.
+	hint	@pause
+	and	scr2=scr1,mask;;	// mask of writecount bits
+	cmp.ne	p6,p0=zeroval,scr2
+(p6)	br.cond.sptk 5b
+	
+	tbit.nz	p8,p0=scr1,DEADLOCKBIT;;// Test for DEADLOCK
+
+	mov psr.l=psrsave;;		// Reenable IC
+	srlz.i;;
+	////////////// END   PHYSICAL MODE ////////////////////
+
+(p8)	add	r8=1,r8
+(p8)	br.cond.spnt 1b;;		// Repeat if DEADLOCK occurred.
+
+	br.ret.sptk	rp
+	.endp sn2_ptc_deadlock_recovery_core
diff --git a/kernel/arch/ia64/sn/kernel/sn2/sn2_smp.c b/kernel/arch/ia64/sn/kernel/sn2/sn2_smp.c
new file mode 100644
index 000000000..f9c8d9fc5
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/sn2_smp.c
@@ -0,0 +1,572 @@
+/*
+ * SN2 Platform specific SMP Support
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/spinlock.h>
+#include <linux/threads.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/mmzone.h>
+#include <linux/module.h>
+#include <linux/bitops.h>
+#include <linux/nodemask.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+
+#include <asm/processor.h>
+#include <asm/irq.h>
+#include <asm/sal.h>
+#include <asm/delay.h>
+#include <asm/io.h>
+#include <asm/smp.h>
+#include <asm/tlb.h>
+#include <asm/numa.h>
+#include <asm/hw_irq.h>
+#include <asm/current.h>
+#include <asm/sn/sn_cpuid.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/shub_mmr.h>
+#include <asm/sn/nodepda.h>
+#include <asm/sn/rw_mmr.h>
+#include <asm/sn/sn_feature_sets.h>
+
+DEFINE_PER_CPU(struct ptc_stats, ptcstats);
+DECLARE_PER_CPU(struct ptc_stats, ptcstats);
+
+static  __cacheline_aligned DEFINE_SPINLOCK(sn2_global_ptc_lock);
+
+/* 0 = old algorithm (no IPI flushes), 1 = ipi deadlock flush, 2 = ipi instead of SHUB ptc, >2 = always ipi */
+static int sn2_flush_opt = 0;
+
+extern unsigned long
+sn2_ptc_deadlock_recovery_core(volatile unsigned long *, unsigned long,
+			       volatile unsigned long *, unsigned long,
+			       volatile unsigned long *, unsigned long);
+void
+sn2_ptc_deadlock_recovery(short *, short, short, int,
+			  volatile unsigned long *, unsigned long,
+			  volatile unsigned long *, unsigned long);
+
+/*
+ * Note: some is the following is captured here to make degugging easier
+ * (the macros make more sense if you see the debug patch - not posted)
+ */
+#define sn2_ptctest	0
+#define local_node_uses_ptc_ga(sh1)	((sh1) ? 1 : 0)
+#define max_active_pio(sh1)		((sh1) ? 32 : 7)
+#define reset_max_active_on_deadlock()	1
+#define PTC_LOCK(sh1)			((sh1) ? &sn2_global_ptc_lock : &sn_nodepda->ptc_lock)
+
+struct ptc_stats {
+	unsigned long ptc_l;
+	unsigned long change_rid;
+	unsigned long shub_ptc_flushes;
+	unsigned long nodes_flushed;
+	unsigned long deadlocks;
+	unsigned long deadlocks2;
+	unsigned long lock_itc_clocks;
+	unsigned long shub_itc_clocks;
+	unsigned long shub_itc_clocks_max;
+	unsigned long shub_ptc_flushes_not_my_mm;
+	unsigned long shub_ipi_flushes;
+	unsigned long shub_ipi_flushes_itc_clocks;
+};
+
+#define sn2_ptctest	0
+
+static inline unsigned long wait_piowc(void)
+{
+	volatile unsigned long *piows;
+	unsigned long zeroval, ws;
+
+	piows = pda->pio_write_status_addr;
+	zeroval = pda->pio_write_status_val;
+	do {
+		cpu_relax();
+	} while (((ws = *piows) & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK) != zeroval);
+	return (ws & SH_PIO_WRITE_STATUS_WRITE_DEADLOCK_MASK) != 0;
+}
+
+/**
+ * sn_migrate - SN-specific task migration actions
+ * @task: Task being migrated to new CPU
+ *
+ * SN2 PIO writes from separate CPUs are not guaranteed to arrive in order.
+ * Context switching user threads which have memory-mapped MMIO may cause
+ * PIOs to issue from separate CPUs, thus the PIO writes must be drained
+ * from the previous CPU's Shub before execution resumes on the new CPU.
+ */
+void sn_migrate(struct task_struct *task)
+{
+	pda_t *last_pda = pdacpu(task_thread_info(task)->last_cpu);
+	volatile unsigned long *adr = last_pda->pio_write_status_addr;
+	unsigned long val = last_pda->pio_write_status_val;
+
+	/* Drain PIO writes from old CPU's Shub */
+	while (unlikely((*adr & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK)
+			!= val))
+		cpu_relax();
+}
+
+void sn_tlb_migrate_finish(struct mm_struct *mm)
+{
+	/* flush_tlb_mm is inefficient if more than 1 users of mm */
+	if (mm == current->mm && mm && atomic_read(&mm->mm_users) == 1)
+		flush_tlb_mm(mm);
+}
+
+static void
+sn2_ipi_flush_all_tlb(struct mm_struct *mm)
+{
+	unsigned long itc;
+
+	itc = ia64_get_itc();
+	smp_flush_tlb_cpumask(*mm_cpumask(mm));
+	itc = ia64_get_itc() - itc;
+	__this_cpu_add(ptcstats.shub_ipi_flushes_itc_clocks, itc);
+	__this_cpu_inc(ptcstats.shub_ipi_flushes);
+}
+
+/**
+ * sn2_global_tlb_purge - globally purge translation cache of virtual address range
+ * @mm: mm_struct containing virtual address range
+ * @start: start of virtual address range
+ * @end: end of virtual address range
+ * @nbits: specifies number of bytes to purge per instruction (num = 1<<(nbits & 0xfc))
+ *
+ * Purges the translation caches of all processors of the given virtual address
+ * range.
+ *
+ * Note:
+ * 	- cpu_vm_mask is a bit mask that indicates which cpus have loaded the context.
+ * 	- cpu_vm_mask is converted into a nodemask of the nodes containing the
+ * 	  cpus in cpu_vm_mask.
+ *	- if only one bit is set in cpu_vm_mask & it is the current cpu & the
+ *	  process is purging its own virtual address range, then only the
+ *	  local TLB needs to be flushed. This flushing can be done using
+ *	  ptc.l. This is the common case & avoids the global spinlock.
+ *	- if multiple cpus have loaded the context, then flushing has to be
+ *	  done with ptc.g/MMRs under protection of the global ptc_lock.
+ */
+
+void
+sn2_global_tlb_purge(struct mm_struct *mm, unsigned long start,
+		     unsigned long end, unsigned long nbits)
+{
+	int i, ibegin, shub1, cnode, mynasid, cpu, lcpu = 0, nasid;
+	int mymm = (mm == current->active_mm && mm == current->mm);
+	int use_cpu_ptcga;
+	volatile unsigned long *ptc0, *ptc1;
+	unsigned long itc, itc2, flags, data0 = 0, data1 = 0, rr_value, old_rr = 0;
+	short nasids[MAX_NUMNODES], nix;
+	nodemask_t nodes_flushed;
+	int active, max_active, deadlock, flush_opt = sn2_flush_opt;
+
+	if (flush_opt > 2) {
+		sn2_ipi_flush_all_tlb(mm);
+		return;
+	}
+
+	nodes_clear(nodes_flushed);
+	i = 0;
+
+	for_each_cpu(cpu, mm_cpumask(mm)) {
+		cnode = cpu_to_node(cpu);
+		node_set(cnode, nodes_flushed);
+		lcpu = cpu;
+		i++;
+	}
+
+	if (i == 0)
+		return;
+
+	preempt_disable();
+
+	if (likely(i == 1 && lcpu == smp_processor_id() && mymm)) {
+		do {
+			ia64_ptcl(start, nbits << 2);
+			start += (1UL << nbits);
+		} while (start < end);
+		ia64_srlz_i();
+		__this_cpu_inc(ptcstats.ptc_l);
+		preempt_enable();
+		return;
+	}
+
+	if (atomic_read(&mm->mm_users) == 1 && mymm) {
+		flush_tlb_mm(mm);
+		__this_cpu_inc(ptcstats.change_rid);
+		preempt_enable();
+		return;
+	}
+
+	if (flush_opt == 2) {
+		sn2_ipi_flush_all_tlb(mm);
+		preempt_enable();
+		return;
+	}
+
+	itc = ia64_get_itc();
+	nix = 0;
+	for_each_node_mask(cnode, nodes_flushed)
+		nasids[nix++] = cnodeid_to_nasid(cnode);
+
+	rr_value = (mm->context << 3) | REGION_NUMBER(start);
+
+	shub1 = is_shub1();
+	if (shub1) {
+		data0 = (1UL << SH1_PTC_0_A_SHFT) |
+		    	(nbits << SH1_PTC_0_PS_SHFT) |
+			(rr_value << SH1_PTC_0_RID_SHFT) |
+		    	(1UL << SH1_PTC_0_START_SHFT);
+		ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_0);
+		ptc1 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_1);
+	} else {
+		data0 = (1UL << SH2_PTC_A_SHFT) |
+			(nbits << SH2_PTC_PS_SHFT) |
+		    	(1UL << SH2_PTC_START_SHFT);
+		ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH2_PTC + 
+			(rr_value << SH2_PTC_RID_SHFT));
+		ptc1 = NULL;
+	}
+	
+
+	mynasid = get_nasid();
+	use_cpu_ptcga = local_node_uses_ptc_ga(shub1);
+	max_active = max_active_pio(shub1);
+
+	itc = ia64_get_itc();
+	spin_lock_irqsave(PTC_LOCK(shub1), flags);
+	itc2 = ia64_get_itc();
+
+	__this_cpu_add(ptcstats.lock_itc_clocks, itc2 - itc);
+	__this_cpu_inc(ptcstats.shub_ptc_flushes);
+	__this_cpu_add(ptcstats.nodes_flushed, nix);
+	if (!mymm)
+		 __this_cpu_inc(ptcstats.shub_ptc_flushes_not_my_mm);
+
+	if (use_cpu_ptcga && !mymm) {
+		old_rr = ia64_get_rr(start);
+		ia64_set_rr(start, (old_rr & 0xff) | (rr_value << 8));
+		ia64_srlz_d();
+	}
+
+	wait_piowc();
+	do {
+		if (shub1)
+			data1 = start | (1UL << SH1_PTC_1_START_SHFT);
+		else
+			data0 = (data0 & ~SH2_PTC_ADDR_MASK) | (start & SH2_PTC_ADDR_MASK);
+		deadlock = 0;
+		active = 0;
+		for (ibegin = 0, i = 0; i < nix; i++) {
+			nasid = nasids[i];
+			if (use_cpu_ptcga && unlikely(nasid == mynasid)) {
+				ia64_ptcga(start, nbits << 2);
+				ia64_srlz_i();
+			} else {
+				ptc0 = CHANGE_NASID(nasid, ptc0);
+				if (ptc1)
+					ptc1 = CHANGE_NASID(nasid, ptc1);
+				pio_atomic_phys_write_mmrs(ptc0, data0, ptc1, data1);
+				active++;
+			}
+			if (active >= max_active || i == (nix - 1)) {
+				if ((deadlock = wait_piowc())) {
+					if (flush_opt == 1)
+						goto done;
+					sn2_ptc_deadlock_recovery(nasids, ibegin, i, mynasid, ptc0, data0, ptc1, data1);
+					if (reset_max_active_on_deadlock())
+						max_active = 1;
+				}
+				active = 0;
+				ibegin = i + 1;
+			}
+		}
+		start += (1UL << nbits);
+	} while (start < end);
+
+done:
+	itc2 = ia64_get_itc() - itc2;
+	__this_cpu_add(ptcstats.shub_itc_clocks, itc2);
+	if (itc2 > __this_cpu_read(ptcstats.shub_itc_clocks_max))
+		__this_cpu_write(ptcstats.shub_itc_clocks_max, itc2);
+
+	if (old_rr) {
+		ia64_set_rr(start, old_rr);
+		ia64_srlz_d();
+	}
+
+	spin_unlock_irqrestore(PTC_LOCK(shub1), flags);
+
+	if (flush_opt == 1 && deadlock) {
+		__this_cpu_inc(ptcstats.deadlocks);
+		sn2_ipi_flush_all_tlb(mm);
+	}
+
+	preempt_enable();
+}
+
+/*
+ * sn2_ptc_deadlock_recovery
+ *
+ * Recover from PTC deadlocks conditions. Recovery requires stepping thru each 
+ * TLB flush transaction.  The recovery sequence is somewhat tricky & is
+ * coded in assembly language.
+ */
+
+void
+sn2_ptc_deadlock_recovery(short *nasids, short ib, short ie, int mynasid,
+			  volatile unsigned long *ptc0, unsigned long data0,
+			  volatile unsigned long *ptc1, unsigned long data1)
+{
+	short nasid, i;
+	unsigned long *piows, zeroval, n;
+
+	__this_cpu_inc(ptcstats.deadlocks);
+
+	piows = (unsigned long *) pda->pio_write_status_addr;
+	zeroval = pda->pio_write_status_val;
+
+
+	for (i=ib; i <= ie; i++) {
+		nasid = nasids[i];
+		if (local_node_uses_ptc_ga(is_shub1()) && nasid == mynasid)
+			continue;
+		ptc0 = CHANGE_NASID(nasid, ptc0);
+		if (ptc1)
+			ptc1 = CHANGE_NASID(nasid, ptc1);
+
+		n = sn2_ptc_deadlock_recovery_core(ptc0, data0, ptc1, data1, piows, zeroval);
+		__this_cpu_add(ptcstats.deadlocks2, n);
+	}
+
+}
+
+/**
+ * sn_send_IPI_phys - send an IPI to a Nasid and slice
+ * @nasid: nasid to receive the interrupt (may be outside partition)
+ * @physid: physical cpuid to receive the interrupt.
+ * @vector: command to send
+ * @delivery_mode: delivery mechanism
+ *
+ * Sends an IPI (interprocessor interrupt) to the processor specified by
+ * @physid
+ *
+ * @delivery_mode can be one of the following
+ *
+ * %IA64_IPI_DM_INT - pend an interrupt
+ * %IA64_IPI_DM_PMI - pend a PMI
+ * %IA64_IPI_DM_NMI - pend an NMI
+ * %IA64_IPI_DM_INIT - pend an INIT interrupt
+ */
+void sn_send_IPI_phys(int nasid, long physid, int vector, int delivery_mode)
+{
+	long val;
+	unsigned long flags = 0;
+	volatile long *p;
+
+	p = (long *)GLOBAL_MMR_PHYS_ADDR(nasid, SH_IPI_INT);
+	val = (1UL << SH_IPI_INT_SEND_SHFT) |
+	    (physid << SH_IPI_INT_PID_SHFT) |
+	    ((long)delivery_mode << SH_IPI_INT_TYPE_SHFT) |
+	    ((long)vector << SH_IPI_INT_IDX_SHFT) |
+	    (0x000feeUL << SH_IPI_INT_BASE_SHFT);
+
+	mb();
+	if (enable_shub_wars_1_1()) {
+		spin_lock_irqsave(&sn2_global_ptc_lock, flags);
+	}
+	pio_phys_write_mmr(p, val);
+	if (enable_shub_wars_1_1()) {
+		wait_piowc();
+		spin_unlock_irqrestore(&sn2_global_ptc_lock, flags);
+	}
+
+}
+
+EXPORT_SYMBOL(sn_send_IPI_phys);
+
+/**
+ * sn2_send_IPI - send an IPI to a processor
+ * @cpuid: target of the IPI
+ * @vector: command to send
+ * @delivery_mode: delivery mechanism
+ * @redirect: redirect the IPI?
+ *
+ * Sends an IPI (InterProcessor Interrupt) to the processor specified by
+ * @cpuid.  @vector specifies the command to send, while @delivery_mode can 
+ * be one of the following
+ *
+ * %IA64_IPI_DM_INT - pend an interrupt
+ * %IA64_IPI_DM_PMI - pend a PMI
+ * %IA64_IPI_DM_NMI - pend an NMI
+ * %IA64_IPI_DM_INIT - pend an INIT interrupt
+ */
+void sn2_send_IPI(int cpuid, int vector, int delivery_mode, int redirect)
+{
+	long physid;
+	int nasid;
+
+	physid = cpu_physical_id(cpuid);
+	nasid = cpuid_to_nasid(cpuid);
+
+	/* the following is used only when starting cpus at boot time */
+	if (unlikely(nasid == -1))
+		ia64_sn_get_sapic_info(physid, &nasid, NULL, NULL);
+
+	sn_send_IPI_phys(nasid, physid, vector, delivery_mode);
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+/**
+ * sn_cpu_disable_allowed - Determine if a CPU can be disabled.
+ * @cpu - CPU that is requested to be disabled.
+ *
+ * CPU disable is only allowed on SHub2 systems running with a PROM
+ * that supports CPU disable. It is not permitted to disable the boot processor.
+ */
+bool sn_cpu_disable_allowed(int cpu)
+{
+	if (is_shub2() && sn_prom_feature_available(PRF_CPU_DISABLE_SUPPORT)) {
+		if (cpu != 0)
+			return true;
+		else
+			printk(KERN_WARNING
+			      "Disabling the boot processor is not allowed.\n");
+
+	} else
+		printk(KERN_WARNING
+		       "CPU disable is not supported on this system.\n");
+
+	return false;
+}
+#endif /* CONFIG_HOTPLUG_CPU */
+
+#ifdef CONFIG_PROC_FS
+
+#define PTC_BASENAME	"sgi_sn/ptc_statistics"
+
+static void *sn2_ptc_seq_start(struct seq_file *file, loff_t * offset)
+{
+	if (*offset < nr_cpu_ids)
+		return offset;
+	return NULL;
+}
+
+static void *sn2_ptc_seq_next(struct seq_file *file, void *data, loff_t * offset)
+{
+	(*offset)++;
+	if (*offset < nr_cpu_ids)
+		return offset;
+	return NULL;
+}
+
+static void sn2_ptc_seq_stop(struct seq_file *file, void *data)
+{
+}
+
+static int sn2_ptc_seq_show(struct seq_file *file, void *data)
+{
+	struct ptc_stats *stat;
+	int cpu;
+
+	cpu = *(loff_t *) data;
+
+	if (!cpu) {
+		seq_printf(file,
+			   "# cpu ptc_l newrid ptc_flushes nodes_flushed deadlocks lock_nsec shub_nsec shub_nsec_max not_my_mm deadlock2 ipi_fluches ipi_nsec\n");
+		seq_printf(file, "# ptctest %d, flushopt %d\n", sn2_ptctest, sn2_flush_opt);
+	}
+
+	if (cpu < nr_cpu_ids && cpu_online(cpu)) {
+		stat = &per_cpu(ptcstats, cpu);
+		seq_printf(file, "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n", cpu, stat->ptc_l,
+				stat->change_rid, stat->shub_ptc_flushes, stat->nodes_flushed,
+				stat->deadlocks,
+				1000 * stat->lock_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
+				1000 * stat->shub_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
+				1000 * stat->shub_itc_clocks_max / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
+				stat->shub_ptc_flushes_not_my_mm,
+				stat->deadlocks2,
+				stat->shub_ipi_flushes,
+				1000 * stat->shub_ipi_flushes_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec);
+	}
+	return 0;
+}
+
+static ssize_t sn2_ptc_proc_write(struct file *file, const char __user *user, size_t count, loff_t *data)
+{
+	int cpu;
+	char optstr[64];
+
+	if (count == 0 || count > sizeof(optstr))
+		return -EINVAL;
+	if (copy_from_user(optstr, user, count))
+		return -EFAULT;
+	optstr[count - 1] = '\0';
+	sn2_flush_opt = simple_strtoul(optstr, NULL, 0);
+
+	for_each_online_cpu(cpu)
+		memset(&per_cpu(ptcstats, cpu), 0, sizeof(struct ptc_stats));
+
+	return count;
+}
+
+static const struct seq_operations sn2_ptc_seq_ops = {
+	.start = sn2_ptc_seq_start,
+	.next = sn2_ptc_seq_next,
+	.stop = sn2_ptc_seq_stop,
+	.show = sn2_ptc_seq_show
+};
+
+static int sn2_ptc_proc_open(struct inode *inode, struct file *file)
+{
+	return seq_open(file, &sn2_ptc_seq_ops);
+}
+
+static const struct file_operations proc_sn2_ptc_operations = {
+	.open = sn2_ptc_proc_open,
+	.read = seq_read,
+	.write = sn2_ptc_proc_write,
+	.llseek = seq_lseek,
+	.release = seq_release,
+};
+
+static struct proc_dir_entry *proc_sn2_ptc;
+
+static int __init sn2_ptc_init(void)
+{
+	if (!ia64_platform_is("sn2"))
+		return 0;
+
+	proc_sn2_ptc = proc_create(PTC_BASENAME, 0444,
+				   NULL, &proc_sn2_ptc_operations);
+	if (!proc_sn2_ptc) {
+		printk(KERN_ERR "unable to create %s proc entry", PTC_BASENAME);
+		return -EINVAL;
+	}
+	spin_lock_init(&sn2_global_ptc_lock);
+	return 0;
+}
+
+static void __exit sn2_ptc_exit(void)
+{
+	remove_proc_entry(PTC_BASENAME, NULL);
+}
+
+module_init(sn2_ptc_init);
+module_exit(sn2_ptc_exit);
+#endif /* CONFIG_PROC_FS */
+
diff --git a/kernel/arch/ia64/sn/kernel/sn2/sn_hwperf.c b/kernel/arch/ia64/sn/kernel/sn2/sn_hwperf.c
new file mode 100644
index 000000000..b9992571c
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/sn_hwperf.c
@@ -0,0 +1,1003 @@
+/* 
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2004-2006 Silicon Graphics, Inc. All rights reserved.
+ *
+ * SGI Altix topology and hardware performance monitoring API.
+ * Mark Goodwin <markgw@sgi.com>. 
+ *
+ * Creates /proc/sgi_sn/sn_topology (read-only) to export
+ * info about Altix nodes, routers, CPUs and NumaLink
+ * interconnection/topology.
+ *
+ * Also creates a dynamic misc device named "sn_hwperf"
+ * that supports an ioctl interface to call down into SAL
+ * to discover hw objects, topology and to read/write
+ * memory mapped registers, e.g. for performance monitoring.
+ * The "sn_hwperf" device is registered only after the procfs
+ * file is first opened, i.e. only if/when it's needed. 
+ *
+ * This API is used by SGI Performance Co-Pilot and other
+ * tools, see http://oss.sgi.com/projects/pcp
+ */
+
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/export.h>
+#include <linux/vmalloc.h>
+#include <linux/seq_file.h>
+#include <linux/miscdevice.h>
+#include <linux/utsname.h>
+#include <linux/cpumask.h>
+#include <linux/nodemask.h>
+#include <linux/smp.h>
+#include <linux/mutex.h>
+
+#include <asm/processor.h>
+#include <asm/topology.h>
+#include <asm/uaccess.h>
+#include <asm/sal.h>
+#include <asm/sn/io.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/module.h>
+#include <asm/sn/geo.h>
+#include <asm/sn/sn2/sn_hwperf.h>
+#include <asm/sn/addrs.h>
+
+static void *sn_hwperf_salheap = NULL;
+static int sn_hwperf_obj_cnt = 0;
+static nasid_t sn_hwperf_master_nasid = INVALID_NASID;
+static int sn_hwperf_init(void);
+static DEFINE_MUTEX(sn_hwperf_init_mutex);
+
+#define cnode_possible(n)	((n) < num_cnodes)
+
+static int sn_hwperf_enum_objects(int *nobj, struct sn_hwperf_object_info **ret)
+{
+	int e;
+	u64 sz;
+	struct sn_hwperf_object_info *objbuf = NULL;
+
+	if ((e = sn_hwperf_init()) < 0) {
+		printk(KERN_ERR "sn_hwperf_init failed: err %d\n", e);
+		goto out;
+	}
+
+	sz = sn_hwperf_obj_cnt * sizeof(struct sn_hwperf_object_info);
+	objbuf = vmalloc(sz);
+	if (objbuf == NULL) {
+		printk("sn_hwperf_enum_objects: vmalloc(%d) failed\n", (int)sz);
+		e = -ENOMEM;
+		goto out;
+	}
+
+	e = ia64_sn_hwperf_op(sn_hwperf_master_nasid, SN_HWPERF_ENUM_OBJECTS,
+		0, sz, (u64) objbuf, 0, 0, NULL);
+	if (e != SN_HWPERF_OP_OK) {
+		e = -EINVAL;
+		vfree(objbuf);
+	}
+
+out:
+	*nobj = sn_hwperf_obj_cnt;
+	*ret = objbuf;
+	return e;
+}
+
+static int sn_hwperf_location_to_bpos(char *location,
+	int *rack, int *bay, int *slot, int *slab)
+{
+	char type;
+
+	/* first scan for an old style geoid string */
+	if (sscanf(location, "%03d%c%02d#%d",
+		rack, &type, bay, slab) == 4)
+		*slot = 0; 
+	else /* scan for a new bladed geoid string */
+	if (sscanf(location, "%03d%c%02d^%02d#%d",
+		rack, &type, bay, slot, slab) != 5)
+		return -1; 
+	/* success */
+	return 0;
+}
+
+static int sn_hwperf_geoid_to_cnode(char *location)
+{
+	int cnode;
+	geoid_t geoid;
+	moduleid_t module_id;
+	int rack, bay, slot, slab;
+	int this_rack, this_bay, this_slot, this_slab;
+
+	if (sn_hwperf_location_to_bpos(location, &rack, &bay, &slot, &slab))
+		return -1;
+
+	/*
+	 * FIXME: replace with cleaner for_each_XXX macro which addresses
+	 * both compute and IO nodes once ACPI3.0 is available.
+	 */
+	for (cnode = 0; cnode < num_cnodes; cnode++) {
+		geoid = cnodeid_get_geoid(cnode);
+		module_id = geo_module(geoid);
+		this_rack = MODULE_GET_RACK(module_id);
+		this_bay = MODULE_GET_BPOS(module_id);
+		this_slot = geo_slot(geoid);
+		this_slab = geo_slab(geoid);
+		if (rack == this_rack && bay == this_bay &&
+			slot == this_slot && slab == this_slab) {
+			break;
+		}
+	}
+
+	return cnode_possible(cnode) ? cnode : -1;
+}
+
+static int sn_hwperf_obj_to_cnode(struct sn_hwperf_object_info * obj)
+{
+	if (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj))
+		BUG();
+	if (SN_HWPERF_FOREIGN(obj))
+		return -1;
+	return sn_hwperf_geoid_to_cnode(obj->location);
+}
+
+static int sn_hwperf_generic_ordinal(struct sn_hwperf_object_info *obj,
+				struct sn_hwperf_object_info *objs)
+{
+	int ordinal;
+	struct sn_hwperf_object_info *p;
+
+	for (ordinal=0, p=objs; p != obj; p++) {
+		if (SN_HWPERF_FOREIGN(p))
+			continue;
+		if (SN_HWPERF_SAME_OBJTYPE(p, obj))
+			ordinal++;
+	}
+
+	return ordinal;
+}
+
+static const char *slabname_node =	"node"; /* SHub asic */
+static const char *slabname_ionode =	"ionode"; /* TIO asic */
+static const char *slabname_router =	"router"; /* NL3R or NL4R */
+static const char *slabname_other =	"other"; /* unknown asic */
+
+static const char *sn_hwperf_get_slabname(struct sn_hwperf_object_info *obj,
+			struct sn_hwperf_object_info *objs, int *ordinal)
+{
+	int isnode;
+	const char *slabname = slabname_other;
+
+	if ((isnode = SN_HWPERF_IS_NODE(obj)) || SN_HWPERF_IS_IONODE(obj)) {
+	    	slabname = isnode ? slabname_node : slabname_ionode;
+		*ordinal = sn_hwperf_obj_to_cnode(obj);
+	}
+	else {
+		*ordinal = sn_hwperf_generic_ordinal(obj, objs);
+		if (SN_HWPERF_IS_ROUTER(obj))
+			slabname = slabname_router;
+	}
+
+	return slabname;
+}
+
+static void print_pci_topology(struct seq_file *s)
+{
+	char *p;
+	size_t sz;
+	int e;
+
+	for (sz = PAGE_SIZE; sz < 16 * PAGE_SIZE; sz += PAGE_SIZE) {
+		if (!(p = kmalloc(sz, GFP_KERNEL)))
+			break;
+		e = ia64_sn_ioif_get_pci_topology(__pa(p), sz);
+		if (e == SALRET_OK)
+			seq_puts(s, p);
+		kfree(p);
+		if (e == SALRET_OK || e == SALRET_NOT_IMPLEMENTED)
+			break;
+	}
+}
+
+static inline int sn_hwperf_has_cpus(cnodeid_t node)
+{
+	return node < MAX_NUMNODES && node_online(node) && nr_cpus_node(node);
+}
+
+static inline int sn_hwperf_has_mem(cnodeid_t node)
+{
+	return node < MAX_NUMNODES && node_online(node) && NODE_DATA(node)->node_present_pages;
+}
+
+static struct sn_hwperf_object_info *
+sn_hwperf_findobj_id(struct sn_hwperf_object_info *objbuf,
+	int nobj, int id)
+{
+	int i;
+	struct sn_hwperf_object_info *p = objbuf;
+
+	for (i=0; i < nobj; i++, p++) {
+		if (p->id == id)
+			return p;
+	}
+
+	return NULL;
+
+}
+
+static int sn_hwperf_get_nearest_node_objdata(struct sn_hwperf_object_info *objbuf,
+	int nobj, cnodeid_t node, cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
+{
+	int e;
+	struct sn_hwperf_object_info *nodeobj = NULL;
+	struct sn_hwperf_object_info *op;
+	struct sn_hwperf_object_info *dest;
+	struct sn_hwperf_object_info *router;
+	struct sn_hwperf_port_info ptdata[16];
+	int sz, i, j;
+	cnodeid_t c;
+	int found_mem = 0;
+	int found_cpu = 0;
+
+	if (!cnode_possible(node))
+		return -EINVAL;
+
+	if (sn_hwperf_has_cpus(node)) {
+		if (near_cpu_node)
+			*near_cpu_node = node;
+		found_cpu++;
+	}
+
+	if (sn_hwperf_has_mem(node)) {
+		if (near_mem_node)
+			*near_mem_node = node;
+		found_mem++;
+	}
+
+	if (found_cpu && found_mem)
+		return 0; /* trivially successful */
+
+	/* find the argument node object */
+	for (i=0, op=objbuf; i < nobj; i++, op++) {
+		if (!SN_HWPERF_IS_NODE(op) && !SN_HWPERF_IS_IONODE(op))
+			continue;
+		if (node == sn_hwperf_obj_to_cnode(op)) {
+			nodeobj = op;
+			break;
+		}
+	}
+	if (!nodeobj) {
+		e = -ENOENT;
+		goto err;
+	}
+
+	/* get it's interconnect topology */
+	sz = op->ports * sizeof(struct sn_hwperf_port_info);
+	BUG_ON(sz > sizeof(ptdata));
+	e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+			      SN_HWPERF_ENUM_PORTS, nodeobj->id, sz,
+			      (u64)&ptdata, 0, 0, NULL);
+	if (e != SN_HWPERF_OP_OK) {
+		e = -EINVAL;
+		goto err;
+	}
+
+	/* find nearest node with cpus and nearest memory */
+	for (router=NULL, j=0; j < op->ports; j++) {
+		dest = sn_hwperf_findobj_id(objbuf, nobj, ptdata[j].conn_id);
+		if (dest && SN_HWPERF_IS_ROUTER(dest))
+			router = dest;
+		if (!dest || SN_HWPERF_FOREIGN(dest) ||
+		    !SN_HWPERF_IS_NODE(dest) || SN_HWPERF_IS_IONODE(dest)) {
+			continue;
+		}
+		c = sn_hwperf_obj_to_cnode(dest);
+		if (!found_cpu && sn_hwperf_has_cpus(c)) {
+			if (near_cpu_node)
+				*near_cpu_node = c;
+			found_cpu++;
+		}
+		if (!found_mem && sn_hwperf_has_mem(c)) {
+			if (near_mem_node)
+				*near_mem_node = c;
+			found_mem++;
+		}
+	}
+
+	if (router && (!found_cpu || !found_mem)) {
+		/* search for a node connected to the same router */
+		sz = router->ports * sizeof(struct sn_hwperf_port_info);
+		BUG_ON(sz > sizeof(ptdata));
+		e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+				      SN_HWPERF_ENUM_PORTS, router->id, sz,
+				      (u64)&ptdata, 0, 0, NULL);
+		if (e != SN_HWPERF_OP_OK) {
+			e = -EINVAL;
+			goto err;
+		}
+		for (j=0; j < router->ports; j++) {
+			dest = sn_hwperf_findobj_id(objbuf, nobj,
+				ptdata[j].conn_id);
+			if (!dest || dest->id == node ||
+			    SN_HWPERF_FOREIGN(dest) ||
+			    !SN_HWPERF_IS_NODE(dest) ||
+			    SN_HWPERF_IS_IONODE(dest)) {
+				continue;
+			}
+			c = sn_hwperf_obj_to_cnode(dest);
+			if (!found_cpu && sn_hwperf_has_cpus(c)) {
+				if (near_cpu_node)
+					*near_cpu_node = c;
+				found_cpu++;
+			}
+			if (!found_mem && sn_hwperf_has_mem(c)) {
+				if (near_mem_node)
+					*near_mem_node = c;
+				found_mem++;
+			}
+			if (found_cpu && found_mem)
+				break;
+		}
+	}
+
+	if (!found_cpu || !found_mem) {
+		/* resort to _any_ node with CPUs and memory */
+		for (i=0, op=objbuf; i < nobj; i++, op++) {
+			if (SN_HWPERF_FOREIGN(op) ||
+			    SN_HWPERF_IS_IONODE(op) ||
+			    !SN_HWPERF_IS_NODE(op)) {
+				continue;
+			}
+			c = sn_hwperf_obj_to_cnode(op);
+			if (!found_cpu && sn_hwperf_has_cpus(c)) {
+				if (near_cpu_node)
+					*near_cpu_node = c;
+				found_cpu++;
+			}
+			if (!found_mem && sn_hwperf_has_mem(c)) {
+				if (near_mem_node)
+					*near_mem_node = c;
+				found_mem++;
+			}
+			if (found_cpu && found_mem)
+				break;
+		}
+	}
+
+	if (!found_cpu || !found_mem)
+		e = -ENODATA;
+
+err:
+	return e;
+}
+
+
+static int sn_topology_show(struct seq_file *s, void *d)
+{
+	int sz;
+	int pt;
+	int e = 0;
+	int i;
+	int j;
+	const char *slabname;
+	int ordinal;
+	char slice;
+	struct cpuinfo_ia64 *c;
+	struct sn_hwperf_port_info *ptdata;
+	struct sn_hwperf_object_info *p;
+	struct sn_hwperf_object_info *obj = d;	/* this object */
+	struct sn_hwperf_object_info *objs = s->private; /* all objects */
+	u8 shubtype;
+	u8 system_size;
+	u8 sharing_size;
+	u8 partid;
+	u8 coher;
+	u8 nasid_shift;
+	u8 region_size;
+	u16 nasid_mask;
+	int nasid_msb;
+
+	if (obj == objs) {
+		seq_printf(s, "# sn_topology version 2\n");
+		seq_printf(s, "# objtype ordinal location partition"
+			" [attribute value [, ...]]\n");
+
+		if (ia64_sn_get_sn_info(0,
+			&shubtype, &nasid_mask, &nasid_shift, &system_size,
+			&sharing_size, &partid, &coher, &region_size))
+			BUG();
+		for (nasid_msb=63; nasid_msb > 0; nasid_msb--) {
+			if (((u64)nasid_mask << nasid_shift) & (1ULL << nasid_msb))
+				break;
+		}
+		seq_printf(s, "partition %u %s local "
+			"shubtype %s, "
+			"nasid_mask 0x%016llx, "
+			"nasid_bits %d:%d, "
+			"system_size %d, "
+			"sharing_size %d, "
+			"coherency_domain %d, "
+			"region_size %d\n",
+
+			partid, utsname()->nodename,
+			shubtype ? "shub2" : "shub1", 
+			(u64)nasid_mask << nasid_shift, nasid_msb, nasid_shift,
+			system_size, sharing_size, coher, region_size);
+
+		print_pci_topology(s);
+	}
+
+	if (SN_HWPERF_FOREIGN(obj)) {
+		/* private in another partition: not interesting */
+		return 0;
+	}
+
+	for (i = 0; i < SN_HWPERF_MAXSTRING && obj->name[i]; i++) {
+		if (obj->name[i] == ' ')
+			obj->name[i] = '_';
+	}
+
+	slabname = sn_hwperf_get_slabname(obj, objs, &ordinal);
+	seq_printf(s, "%s %d %s %s asic %s", slabname, ordinal, obj->location,
+		obj->sn_hwp_this_part ? "local" : "shared", obj->name);
+
+	if (ordinal < 0 || (!SN_HWPERF_IS_NODE(obj) && !SN_HWPERF_IS_IONODE(obj)))
+		seq_putc(s, '\n');
+	else {
+		cnodeid_t near_mem = -1;
+		cnodeid_t near_cpu = -1;
+
+		seq_printf(s, ", nasid 0x%x", cnodeid_to_nasid(ordinal));
+
+		if (sn_hwperf_get_nearest_node_objdata(objs, sn_hwperf_obj_cnt,
+			ordinal, &near_mem, &near_cpu) == 0) {
+			seq_printf(s, ", near_mem_nodeid %d, near_cpu_nodeid %d",
+				near_mem, near_cpu);
+		}
+
+		if (!SN_HWPERF_IS_IONODE(obj)) {
+			for_each_online_node(i) {
+				seq_printf(s, i ? ":%d" : ", dist %d",
+					node_distance(ordinal, i));
+			}
+		}
+
+		seq_putc(s, '\n');
+
+		/*
+		 * CPUs on this node, if any
+		 */
+		if (!SN_HWPERF_IS_IONODE(obj)) {
+			for_each_cpu_and(i, cpu_online_mask,
+					 cpumask_of_node(ordinal)) {
+				slice = 'a' + cpuid_to_slice(i);
+				c = cpu_data(i);
+				seq_printf(s, "cpu %d %s%c local"
+					   " freq %luMHz, arch ia64",
+					   i, obj->location, slice,
+					   c->proc_freq / 1000000);
+				for_each_online_cpu(j) {
+					seq_printf(s, j ? ":%d" : ", dist %d",
+						   node_distance(
+						    	cpu_to_node(i),
+						    	cpu_to_node(j)));
+				}
+				seq_putc(s, '\n');
+			}
+		}
+	}
+
+	if (obj->ports) {
+		/*
+		 * numalink ports
+		 */
+		sz = obj->ports * sizeof(struct sn_hwperf_port_info);
+		if ((ptdata = kmalloc(sz, GFP_KERNEL)) == NULL)
+			return -ENOMEM;
+		e = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+				      SN_HWPERF_ENUM_PORTS, obj->id, sz,
+				      (u64) ptdata, 0, 0, NULL);
+		if (e != SN_HWPERF_OP_OK)
+			return -EINVAL;
+		for (ordinal=0, p=objs; p != obj; p++) {
+			if (!SN_HWPERF_FOREIGN(p))
+				ordinal += p->ports;
+		}
+		for (pt = 0; pt < obj->ports; pt++) {
+			for (p = objs, i = 0; i < sn_hwperf_obj_cnt; i++, p++) {
+				if (ptdata[pt].conn_id == p->id) {
+					break;
+				}
+			}
+			seq_printf(s, "numalink %d %s-%d",
+			    ordinal+pt, obj->location, ptdata[pt].port);
+
+			if (i >= sn_hwperf_obj_cnt) {
+				/* no connection */
+				seq_puts(s, " local endpoint disconnected"
+					    ", protocol unknown\n");
+				continue;
+			}
+
+			if (obj->sn_hwp_this_part && p->sn_hwp_this_part)
+				/* both ends local to this partition */
+				seq_puts(s, " local");
+			else if (SN_HWPERF_FOREIGN(p))
+				/* both ends of the link in foreign partiton */
+				seq_puts(s, " foreign");
+			else
+				/* link straddles a partition */
+				seq_puts(s, " shared");
+
+			/*
+			 * Unlikely, but strictly should query the LLP config
+			 * registers because an NL4R can be configured to run
+			 * NL3 protocol, even when not talking to an NL3 router.
+			 * Ditto for node-node.
+			 */
+			seq_printf(s, " endpoint %s-%d, protocol %s\n",
+				p->location, ptdata[pt].conn_port,
+				(SN_HWPERF_IS_NL3ROUTER(obj) ||
+				SN_HWPERF_IS_NL3ROUTER(p)) ?  "LLP3" : "LLP4");
+		}
+		kfree(ptdata);
+	}
+
+	return 0;
+}
+
+static void *sn_topology_start(struct seq_file *s, loff_t * pos)
+{
+	struct sn_hwperf_object_info *objs = s->private;
+
+	if (*pos < sn_hwperf_obj_cnt)
+		return (void *)(objs + *pos);
+
+	return NULL;
+}
+
+static void *sn_topology_next(struct seq_file *s, void *v, loff_t * pos)
+{
+	++*pos;
+	return sn_topology_start(s, pos);
+}
+
+static void sn_topology_stop(struct seq_file *m, void *v)
+{
+	return;
+}
+
+/*
+ * /proc/sgi_sn/sn_topology, read-only using seq_file
+ */
+static const struct seq_operations sn_topology_seq_ops = {
+	.start = sn_topology_start,
+	.next = sn_topology_next,
+	.stop = sn_topology_stop,
+	.show = sn_topology_show
+};
+
+struct sn_hwperf_op_info {
+	u64 op;
+	struct sn_hwperf_ioctl_args *a;
+	void *p;
+	int *v0;
+	int ret;
+};
+
+static void sn_hwperf_call_sal(void *info)
+{
+	struct sn_hwperf_op_info *op_info = info;
+	int r;
+
+	r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op_info->op,
+		      op_info->a->arg, op_info->a->sz,
+		      (u64) op_info->p, 0, 0, op_info->v0);
+	op_info->ret = r;
+}
+
+static int sn_hwperf_op_cpu(struct sn_hwperf_op_info *op_info)
+{
+	u32 cpu;
+	u32 use_ipi;
+	int r = 0;
+	cpumask_t save_allowed;
+	
+	cpu = (op_info->a->arg & SN_HWPERF_ARG_CPU_MASK) >> 32;
+	use_ipi = op_info->a->arg & SN_HWPERF_ARG_USE_IPI_MASK;
+	op_info->a->arg &= SN_HWPERF_ARG_OBJID_MASK;
+
+	if (cpu != SN_HWPERF_ARG_ANY_CPU) {
+		if (cpu >= nr_cpu_ids || !cpu_online(cpu)) {
+			r = -EINVAL;
+			goto out;
+		}
+	}
+
+	if (cpu == SN_HWPERF_ARG_ANY_CPU) {
+		/* don't care which cpu */
+		sn_hwperf_call_sal(op_info);
+	} else if (cpu == get_cpu()) {
+		/* already on correct cpu */
+		sn_hwperf_call_sal(op_info);
+		put_cpu();
+	} else {
+		put_cpu();
+		if (use_ipi) {
+			/* use an interprocessor interrupt to call SAL */
+			smp_call_function_single(cpu, sn_hwperf_call_sal,
+				op_info, 1);
+		}
+		else {
+			/* migrate the task before calling SAL */ 
+			save_allowed = current->cpus_allowed;
+			set_cpus_allowed_ptr(current, cpumask_of(cpu));
+			sn_hwperf_call_sal(op_info);
+			set_cpus_allowed_ptr(current, &save_allowed);
+		}
+	}
+	r = op_info->ret;
+
+out:
+	return r;
+}
+
+/* map SAL hwperf error code to system error code */
+static int sn_hwperf_map_err(int hwperf_err)
+{
+	int e;
+
+	switch(hwperf_err) {
+	case SN_HWPERF_OP_OK:
+		e = 0;
+		break;
+
+	case SN_HWPERF_OP_NOMEM:
+		e = -ENOMEM;
+		break;
+
+	case SN_HWPERF_OP_NO_PERM:
+		e = -EPERM;
+		break;
+
+	case SN_HWPERF_OP_IO_ERROR:
+		e = -EIO;
+		break;
+
+	case SN_HWPERF_OP_BUSY:
+		e = -EBUSY;
+		break;
+
+	case SN_HWPERF_OP_RECONFIGURE:
+		e = -EAGAIN;
+		break;
+
+	case SN_HWPERF_OP_INVAL:
+	default:
+		e = -EINVAL;
+		break;
+	}
+
+	return e;
+}
+
+/*
+ * ioctl for "sn_hwperf" misc device
+ */
+static long sn_hwperf_ioctl(struct file *fp, u32 op, unsigned long arg)
+{
+	struct sn_hwperf_ioctl_args a;
+	struct cpuinfo_ia64 *cdata;
+	struct sn_hwperf_object_info *objs;
+	struct sn_hwperf_object_info *cpuobj;
+	struct sn_hwperf_op_info op_info;
+	void *p = NULL;
+	int nobj;
+	char slice;
+	int node;
+	int r;
+	int v0;
+	int i;
+	int j;
+
+	/* only user requests are allowed here */
+	if ((op & SN_HWPERF_OP_MASK) < 10) {
+		r = -EINVAL;
+		goto error;
+	}
+	r = copy_from_user(&a, (const void __user *)arg,
+		sizeof(struct sn_hwperf_ioctl_args));
+	if (r != 0) {
+		r = -EFAULT;
+		goto error;
+	}
+
+	/*
+	 * Allocate memory to hold a kernel copy of the user buffer. The
+	 * buffer contents are either copied in or out (or both) of user
+	 * space depending on the flags encoded in the requested operation.
+	 */
+	if (a.ptr) {
+		p = vmalloc(a.sz);
+		if (!p) {
+			r = -ENOMEM;
+			goto error;
+		}
+	}
+
+	if (op & SN_HWPERF_OP_MEM_COPYIN) {
+		r = copy_from_user(p, (const void __user *)a.ptr, a.sz);
+		if (r != 0) {
+			r = -EFAULT;
+			goto error;
+		}
+	}
+
+	switch (op) {
+	case SN_HWPERF_GET_CPU_INFO:
+		if (a.sz == sizeof(u64)) {
+			/* special case to get size needed */
+			*(u64 *) p = (u64) num_online_cpus() *
+				sizeof(struct sn_hwperf_object_info);
+		} else
+		if (a.sz < num_online_cpus() * sizeof(struct sn_hwperf_object_info)) {
+			r = -ENOMEM;
+			goto error;
+		} else
+		if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
+			int cpuobj_index = 0;
+
+			memset(p, 0, a.sz);
+			for (i = 0; i < nobj; i++) {
+				if (!SN_HWPERF_IS_NODE(objs + i))
+					continue;
+				node = sn_hwperf_obj_to_cnode(objs + i);
+				for_each_online_cpu(j) {
+					if (node != cpu_to_node(j))
+						continue;
+					cpuobj = (struct sn_hwperf_object_info *) p + cpuobj_index++;
+					slice = 'a' + cpuid_to_slice(j);
+					cdata = cpu_data(j);
+					cpuobj->id = j;
+					snprintf(cpuobj->name,
+						 sizeof(cpuobj->name),
+						 "CPU %luMHz %s",
+						 cdata->proc_freq / 1000000,
+						 cdata->vendor);
+					snprintf(cpuobj->location,
+						 sizeof(cpuobj->location),
+						 "%s%c", objs[i].location,
+						 slice);
+				}
+			}
+
+			vfree(objs);
+		}
+		break;
+
+	case SN_HWPERF_GET_NODE_NASID:
+		if (a.sz != sizeof(u64) ||
+		   (node = a.arg) < 0 || !cnode_possible(node)) {
+			r = -EINVAL;
+			goto error;
+		}
+		*(u64 *)p = (u64)cnodeid_to_nasid(node);
+		break;
+
+	case SN_HWPERF_GET_OBJ_NODE:
+		i = a.arg;
+		if (a.sz != sizeof(u64) || i < 0) {
+			r = -EINVAL;
+			goto error;
+		}
+		if ((r = sn_hwperf_enum_objects(&nobj, &objs)) == 0) {
+			if (i >= nobj) {
+				r = -EINVAL;
+				vfree(objs);
+				goto error;
+			}
+			if (objs[i].id != a.arg) {
+				for (i = 0; i < nobj; i++) {
+					if (objs[i].id == a.arg)
+						break;
+				}
+			}
+			if (i == nobj) {
+				r = -EINVAL;
+				vfree(objs);
+				goto error;
+			}
+
+			if (!SN_HWPERF_IS_NODE(objs + i) &&
+			    !SN_HWPERF_IS_IONODE(objs + i)) {
+			    	r = -ENOENT;
+				vfree(objs);
+				goto error;
+			}
+
+			*(u64 *)p = (u64)sn_hwperf_obj_to_cnode(objs + i);
+			vfree(objs);
+		}
+		break;
+
+	case SN_HWPERF_GET_MMRS:
+	case SN_HWPERF_SET_MMRS:
+	case SN_HWPERF_OBJECT_DISTANCE:
+		op_info.p = p;
+		op_info.a = &a;
+		op_info.v0 = &v0;
+		op_info.op = op;
+		r = sn_hwperf_op_cpu(&op_info);
+		if (r) {
+			r = sn_hwperf_map_err(r);
+			a.v0 = v0;
+			goto error;
+		}
+		break;
+
+	default:
+		/* all other ops are a direct SAL call */
+		r = ia64_sn_hwperf_op(sn_hwperf_master_nasid, op,
+			      a.arg, a.sz, (u64) p, 0, 0, &v0);
+		if (r) {
+			r = sn_hwperf_map_err(r);
+			goto error;
+		}
+		a.v0 = v0;
+		break;
+	}
+
+	if (op & SN_HWPERF_OP_MEM_COPYOUT) {
+		r = copy_to_user((void __user *)a.ptr, p, a.sz);
+		if (r != 0) {
+			r = -EFAULT;
+			goto error;
+		}
+	}
+
+error:
+	vfree(p);
+
+	return r;
+}
+
+static const struct file_operations sn_hwperf_fops = {
+	.unlocked_ioctl = sn_hwperf_ioctl,
+	.llseek = noop_llseek,
+};
+
+static struct miscdevice sn_hwperf_dev = {
+	MISC_DYNAMIC_MINOR,
+	"sn_hwperf",
+	&sn_hwperf_fops
+};
+
+static int sn_hwperf_init(void)
+{
+	u64 v;
+	int salr;
+	int e = 0;
+
+	/* single threaded, once-only initialization */
+	mutex_lock(&sn_hwperf_init_mutex);
+
+	if (sn_hwperf_salheap) {
+		mutex_unlock(&sn_hwperf_init_mutex);
+		return e;
+	}
+
+	/*
+	 * The PROM code needs a fixed reference node. For convenience the
+	 * same node as the console I/O is used.
+	 */
+	sn_hwperf_master_nasid = (nasid_t) ia64_sn_get_console_nasid();
+
+	/*
+	 * Request the needed size and install the PROM scratch area.
+	 * The PROM keeps various tracking bits in this memory area.
+	 */
+	salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+				 (u64) SN_HWPERF_GET_HEAPSIZE, 0,
+				 (u64) sizeof(u64), (u64) &v, 0, 0, NULL);
+	if (salr != SN_HWPERF_OP_OK) {
+		e = -EINVAL;
+		goto out;
+	}
+
+	if ((sn_hwperf_salheap = vmalloc(v)) == NULL) {
+		e = -ENOMEM;
+		goto out;
+	}
+	salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+				 SN_HWPERF_INSTALL_HEAP, 0, v,
+				 (u64) sn_hwperf_salheap, 0, 0, NULL);
+	if (salr != SN_HWPERF_OP_OK) {
+		e = -EINVAL;
+		goto out;
+	}
+
+	salr = ia64_sn_hwperf_op(sn_hwperf_master_nasid,
+				 SN_HWPERF_OBJECT_COUNT, 0,
+				 sizeof(u64), (u64) &v, 0, 0, NULL);
+	if (salr != SN_HWPERF_OP_OK) {
+		e = -EINVAL;
+		goto out;
+	}
+	sn_hwperf_obj_cnt = (int)v;
+
+out:
+	if (e < 0 && sn_hwperf_salheap) {
+		vfree(sn_hwperf_salheap);
+		sn_hwperf_salheap = NULL;
+		sn_hwperf_obj_cnt = 0;
+	}
+	mutex_unlock(&sn_hwperf_init_mutex);
+	return e;
+}
+
+int sn_topology_open(struct inode *inode, struct file *file)
+{
+	int e;
+	struct seq_file *seq;
+	struct sn_hwperf_object_info *objbuf;
+	int nobj;
+
+	if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
+		e = seq_open(file, &sn_topology_seq_ops);
+		seq = file->private_data;
+		seq->private = objbuf;
+	}
+
+	return e;
+}
+
+int sn_topology_release(struct inode *inode, struct file *file)
+{
+	struct seq_file *seq = file->private_data;
+
+	vfree(seq->private);
+	return seq_release(inode, file);
+}
+
+int sn_hwperf_get_nearest_node(cnodeid_t node,
+	cnodeid_t *near_mem_node, cnodeid_t *near_cpu_node)
+{
+	int e;
+	int nobj;
+	struct sn_hwperf_object_info *objbuf;
+
+	if ((e = sn_hwperf_enum_objects(&nobj, &objbuf)) == 0) {
+		e = sn_hwperf_get_nearest_node_objdata(objbuf, nobj,
+			node, near_mem_node, near_cpu_node);
+		vfree(objbuf);
+	}
+
+	return e;
+}
+
+static int sn_hwperf_misc_register_init(void)
+{
+	int e;
+
+	if (!ia64_platform_is("sn2"))
+		return 0;
+
+	sn_hwperf_init();
+
+	/*
+	 * Register a dynamic misc device for hwperf ioctls. Platforms
+	 * supporting hotplug will create /dev/sn_hwperf, else user
+	 * can to look up the minor number in /proc/misc.
+	 */
+	if ((e = misc_register(&sn_hwperf_dev)) != 0) {
+		printk(KERN_ERR "sn_hwperf_misc_register_init: failed to "
+		"register misc device for \"%s\"\n", sn_hwperf_dev.name);
+	}
+
+	return e;
+}
+
+device_initcall(sn_hwperf_misc_register_init); /* after misc_init() */
+EXPORT_SYMBOL(sn_hwperf_get_nearest_node);
diff --git a/kernel/arch/ia64/sn/kernel/sn2/sn_proc_fs.c b/kernel/arch/ia64/sn/kernel/sn2/sn_proc_fs.c
new file mode 100644
index 000000000..7aab87f48
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/sn_proc_fs.c
@@ -0,0 +1,117 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#ifdef CONFIG_PROC_FS
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <asm/uaccess.h>
+#include <asm/sn/sn_sal.h>
+
+static int partition_id_show(struct seq_file *s, void *p)
+{
+	seq_printf(s, "%d\n", sn_partition_id);
+	return 0;
+}
+
+static int partition_id_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, partition_id_show, NULL);
+}
+
+static int system_serial_number_show(struct seq_file *s, void *p)
+{
+	seq_printf(s, "%s\n", sn_system_serial_number());
+	return 0;
+}
+
+static int system_serial_number_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, system_serial_number_show, NULL);
+}
+
+static int licenseID_show(struct seq_file *s, void *p)
+{
+	seq_printf(s, "0x%llx\n", sn_partition_serial_number_val());
+	return 0;
+}
+
+static int licenseID_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, licenseID_show, NULL);
+}
+
+static int coherence_id_show(struct seq_file *s, void *p)
+{
+	seq_printf(s, "%d\n", partition_coherence_id());
+
+	return 0;
+}
+
+static int coherence_id_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, coherence_id_show, NULL);
+}
+
+/* /proc/sgi_sn/sn_topology uses seq_file, see sn_hwperf.c */
+extern int sn_topology_open(struct inode *, struct file *);
+extern int sn_topology_release(struct inode *, struct file *);
+
+static const struct file_operations proc_partition_id_fops = {
+	.open		= partition_id_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static const struct file_operations proc_system_sn_fops = {
+	.open		= system_serial_number_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static const struct file_operations proc_license_id_fops = {
+	.open		= licenseID_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static const struct file_operations proc_coherence_id_fops = {
+	.open		= coherence_id_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static const struct file_operations proc_sn_topo_fops = {
+	.open		= sn_topology_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= sn_topology_release,
+};
+
+void register_sn_procfs(void)
+{
+	static struct proc_dir_entry *sgi_proc_dir = NULL;
+
+	BUG_ON(sgi_proc_dir != NULL);
+	if (!(sgi_proc_dir = proc_mkdir("sgi_sn", NULL)))
+		return;
+
+	proc_create("partition_id", 0444, sgi_proc_dir,
+		    &proc_partition_id_fops);
+	proc_create("system_serial_number", 0444, sgi_proc_dir,
+		    &proc_system_sn_fops);
+	proc_create("licenseID", 0444, sgi_proc_dir, &proc_license_id_fops);
+	proc_create("coherence_id", 0444, sgi_proc_dir,
+		    &proc_coherence_id_fops);
+	proc_create("sn_topology", 0444, sgi_proc_dir, &proc_sn_topo_fops);
+}
+
+#endif /* CONFIG_PROC_FS */
diff --git a/kernel/arch/ia64/sn/kernel/sn2/timer.c b/kernel/arch/ia64/sn/kernel/sn2/timer.c
new file mode 100644
index 000000000..abab8f99e
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/timer.c
@@ -0,0 +1,60 @@
+/*
+ * linux/arch/ia64/sn/kernel/sn2/timer.c
+ *
+ * Copyright (C) 2003 Silicon Graphics, Inc.
+ * Copyright (C) 2003 Hewlett-Packard Co
+ *	David Mosberger <davidm@hpl.hp.com>: updated for new timer-interpolation infrastructure
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/time.h>
+#include <linux/interrupt.h>
+#include <linux/clocksource.h>
+
+#include <asm/hw_irq.h>
+#include <asm/timex.h>
+
+#include <asm/sn/leds.h>
+#include <asm/sn/shub_mmr.h>
+#include <asm/sn/clksupport.h>
+
+extern unsigned long sn_rtc_cycles_per_second;
+
+static cycle_t read_sn2(struct clocksource *cs)
+{
+	return (cycle_t)readq(RTC_COUNTER_ADDR);
+}
+
+static struct clocksource clocksource_sn2 = {
+        .name           = "sn2_rtc",
+        .rating         = 450,
+        .read           = read_sn2,
+        .mask           = (1LL << 55) - 1,
+        .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
+};
+
+/*
+ * sn udelay uses the RTC instead of the ITC because the ITC is not
+ * synchronized across all CPUs, and the thread may migrate to another CPU
+ * if preemption is enabled.
+ */
+static void
+ia64_sn_udelay (unsigned long usecs)
+{
+	unsigned long start = rtc_time();
+	unsigned long end = start +
+			usecs * sn_rtc_cycles_per_second / 1000000;
+
+	while (time_before((unsigned long)rtc_time(), end))
+		cpu_relax();
+}
+
+void __init sn_timer_init(void)
+{
+	clocksource_sn2.archdata.fsys_mmio = RTC_COUNTER_ADDR;
+	clocksource_register_hz(&clocksource_sn2, sn_rtc_cycles_per_second);
+
+	ia64_udelay = &ia64_sn_udelay;
+}
diff --git a/kernel/arch/ia64/sn/kernel/sn2/timer_interrupt.c b/kernel/arch/ia64/sn/kernel/sn2/timer_interrupt.c
new file mode 100644
index 000000000..103d6ea8e
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/sn2/timer_interrupt.c
@@ -0,0 +1,60 @@
+/*
+ *
+ *
+ * Copyright (c) 2005, 2006 Silicon Graphics, Inc.  All Rights Reserved.
+ * 
+ * This program is free software; you can redistribute it and/or modify it 
+ * under the terms of version 2 of the GNU General Public License 
+ * as published by the Free Software Foundation.
+ * 
+ * This program is distributed in the hope that it would be useful, but 
+ * WITHOUT ANY WARRANTY; without even the implied warranty of 
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. 
+ * 
+ * Further, this software is distributed without any warranty that it is 
+ * free of the rightful claim of any third person regarding infringement 
+ * or the like.  Any license provided herein, whether implied or 
+ * otherwise, applies only to this software file.  Patent licenses, if 
+ * any, provided herein do not apply to combinations of this program with 
+ * other software, or any other product whatsoever.
+ * 
+ * You should have received a copy of the GNU General Public 
+ * License along with this program; if not, write the Free Software 
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston MA 02111-1307, USA.
+ * 
+ * For further information regarding this notice, see: 
+ * 
+ * http://oss.sgi.com/projects/GenInfo/NoticeExplan
+ */
+
+#include <linux/interrupt.h>
+#include <asm/sn/pda.h>
+#include <asm/sn/leds.h>
+
+extern void sn_lb_int_war_check(void);
+extern irqreturn_t timer_interrupt(int irq, void *dev_id, struct pt_regs *regs);
+
+#define SN_LB_INT_WAR_INTERVAL 100
+
+void sn_timer_interrupt(int irq, void *dev_id)
+{
+	/* LED blinking */
+	if (!pda->hb_count--) {
+		pda->hb_count = HZ / 2;
+		set_led_bits(pda->hb_state ^=
+			     LED_CPU_HEARTBEAT, LED_CPU_HEARTBEAT);
+	}
+
+	if (is_shub1()) {
+		if (enable_shub_wars_1_1()) {
+			/* Bugfix code for SHUB 1.1 */
+			if (pda->pio_shub_war_cam_addr)
+				*pda->pio_shub_war_cam_addr = 0x8000000000000010UL;
+		}
+		if (pda->sn_lb_int_war_ticks == 0)
+			sn_lb_int_war_check();
+		pda->sn_lb_int_war_ticks++;
+		if (pda->sn_lb_int_war_ticks >= SN_LB_INT_WAR_INTERVAL)
+			pda->sn_lb_int_war_ticks = 0;
+	}
+}
diff --git a/kernel/arch/ia64/sn/kernel/tiocx.c b/kernel/arch/ia64/sn/kernel/tiocx.c
new file mode 100644
index 000000000..e35f6485c
--- /dev/null
+++ b/kernel/arch/ia64/sn/kernel/tiocx.c
@@ -0,0 +1,569 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2005 Silicon Graphics, Inc.  All rights reserved.
+ */
+
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/proc_fs.h>
+#include <linux/capability.h>
+#include <linux/device.h>
+#include <linux/delay.h>
+#include <asm/uaccess.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/io.h>
+#include <asm/sn/types.h>
+#include <asm/sn/shubio.h>
+#include <asm/sn/tiocx.h>
+#include <asm/sn/l1.h>
+#include <asm/sn/module.h>
+#include "tio.h"
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+
+#define CX_DEV_NONE 0
+#define DEVICE_NAME "tiocx"
+#define WIDGET_ID 0
+#define TIOCX_DEBUG 0
+
+#if TIOCX_DEBUG
+#define DBG(fmt...)    printk(KERN_ALERT fmt)
+#else
+#define DBG(fmt...)
+#endif
+
+struct device_attribute dev_attr_cxdev_control;
+
+/**
+ * tiocx_match - Try to match driver id list with device.
+ * @dev: device pointer
+ * @drv: driver pointer
+ *
+ * Returns 1 if match, 0 otherwise.
+ */
+static int tiocx_match(struct device *dev, struct device_driver *drv)
+{
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+	struct cx_drv *cx_drv = to_cx_driver(drv);
+	const struct cx_device_id *ids = cx_drv->id_table;
+
+	if (!ids)
+		return 0;
+
+	while (ids->part_num) {
+		if (ids->part_num == cx_dev->cx_id.part_num)
+			return 1;
+		ids++;
+	}
+	return 0;
+
+}
+
+static int tiocx_uevent(struct device *dev, struct kobj_uevent_env *env)
+{
+	return -ENODEV;
+}
+
+static void tiocx_bus_release(struct device *dev)
+{
+	kfree(to_cx_dev(dev));
+}
+
+/**
+ * cx_device_match - Find cx_device in the id table.
+ * @ids: id table from driver
+ * @cx_device: part/mfg id for the device
+ *
+ */
+static const struct cx_device_id *cx_device_match(const struct cx_device_id
+						  *ids,
+						  struct cx_dev *cx_device)
+{
+	/*
+	 * NOTES: We may want to check for CX_ANY_ID too.
+	 *        Do we want to match against nasid too?
+	 *        CX_DEV_NONE == 0, if the driver tries to register for
+	 *        part/mfg == 0 we should return no-match (NULL) here.
+	 */
+	while (ids->part_num && ids->mfg_num) {
+		if (ids->part_num == cx_device->cx_id.part_num &&
+		    ids->mfg_num == cx_device->cx_id.mfg_num)
+			return ids;
+		ids++;
+	}
+
+	return NULL;
+}
+
+/**
+ * cx_device_probe - Look for matching device.
+ *			Call driver probe routine if found.
+ * @cx_driver: driver table (cx_drv struct) from driver
+ * @cx_device: part/mfg id for the device
+ */
+static int cx_device_probe(struct device *dev)
+{
+	const struct cx_device_id *id;
+	struct cx_drv *cx_drv = to_cx_driver(dev->driver);
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+	int error = 0;
+
+	if (!cx_dev->driver && cx_drv->probe) {
+		id = cx_device_match(cx_drv->id_table, cx_dev);
+		if (id) {
+			if ((error = cx_drv->probe(cx_dev, id)) < 0)
+				return error;
+			else
+				cx_dev->driver = cx_drv;
+		}
+	}
+
+	return error;
+}
+
+/**
+ * cx_driver_remove - Remove driver from device struct.
+ * @dev: device
+ */
+static int cx_driver_remove(struct device *dev)
+{
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+	struct cx_drv *cx_drv = cx_dev->driver;
+	if (cx_drv->remove)
+		cx_drv->remove(cx_dev);
+	cx_dev->driver = NULL;
+	return 0;
+}
+
+struct bus_type tiocx_bus_type = {
+	.name = "tiocx",
+	.match = tiocx_match,
+	.uevent = tiocx_uevent,
+	.probe = cx_device_probe,
+	.remove = cx_driver_remove,
+};
+
+/**
+ * cx_driver_register - Register the driver.
+ * @cx_driver: driver table (cx_drv struct) from driver
+ * 
+ * Called from the driver init routine to register a driver.
+ * The cx_drv struct contains the driver name, a pointer to
+ * a table of part/mfg numbers and a pointer to the driver's
+ * probe/attach routine.
+ */
+int cx_driver_register(struct cx_drv *cx_driver)
+{
+	cx_driver->driver.name = cx_driver->name;
+	cx_driver->driver.bus = &tiocx_bus_type;
+
+	return driver_register(&cx_driver->driver);
+}
+
+/**
+ * cx_driver_unregister - Unregister the driver.
+ * @cx_driver: driver table (cx_drv struct) from driver
+ */
+int cx_driver_unregister(struct cx_drv *cx_driver)
+{
+	driver_unregister(&cx_driver->driver);
+	return 0;
+}
+
+/**
+ * cx_device_register - Register a device.
+ * @nasid: device's nasid
+ * @part_num: device's part number
+ * @mfg_num: device's manufacturer number
+ * @hubdev: hub info associated with this device
+ * @bt: board type of the device
+ *
+ */
+int
+cx_device_register(nasid_t nasid, int part_num, int mfg_num,
+		   struct hubdev_info *hubdev, int bt)
+{
+	struct cx_dev *cx_dev;
+	int r;
+
+	cx_dev = kzalloc(sizeof(struct cx_dev), GFP_KERNEL);
+	DBG("cx_dev= 0x%p\n", cx_dev);
+	if (cx_dev == NULL)
+		return -ENOMEM;
+
+	cx_dev->cx_id.part_num = part_num;
+	cx_dev->cx_id.mfg_num = mfg_num;
+	cx_dev->cx_id.nasid = nasid;
+	cx_dev->hubdev = hubdev;
+	cx_dev->bt = bt;
+
+	cx_dev->dev.parent = NULL;
+	cx_dev->dev.bus = &tiocx_bus_type;
+	cx_dev->dev.release = tiocx_bus_release;
+	dev_set_name(&cx_dev->dev, "%d", cx_dev->cx_id.nasid);
+	r = device_register(&cx_dev->dev);
+	if (r) {
+		kfree(cx_dev);
+		return r;
+	}
+	get_device(&cx_dev->dev);
+
+	device_create_file(&cx_dev->dev, &dev_attr_cxdev_control);
+
+	return 0;
+}
+
+/**
+ * cx_device_unregister - Unregister a device.
+ * @cx_dev: part/mfg id for the device
+ */
+int cx_device_unregister(struct cx_dev *cx_dev)
+{
+	put_device(&cx_dev->dev);
+	device_unregister(&cx_dev->dev);
+	return 0;
+}
+
+/**
+ * cx_device_reload - Reload the device.
+ * @nasid: device's nasid
+ * @part_num: device's part number
+ * @mfg_num: device's manufacturer number
+ *
+ * Remove the device associated with 'nasid' from device list and then
+ * call device-register with the given part/mfg numbers.
+ */
+static int cx_device_reload(struct cx_dev *cx_dev)
+{
+	cx_device_unregister(cx_dev);
+	return cx_device_register(cx_dev->cx_id.nasid, cx_dev->cx_id.part_num,
+				  cx_dev->cx_id.mfg_num, cx_dev->hubdev,
+				  cx_dev->bt);
+}
+
+static inline u64 tiocx_intr_alloc(nasid_t nasid, int widget,
+					u64 sn_irq_info,
+					int req_irq, nasid_t req_nasid,
+					int req_slice)
+{
+	struct ia64_sal_retval rv;
+	rv.status = 0;
+	rv.v0 = 0;
+
+	ia64_sal_oemcall_nolock(&rv, SN_SAL_IOIF_INTERRUPT,
+				SAL_INTR_ALLOC, nasid,
+				widget, sn_irq_info, req_irq,
+				req_nasid, req_slice);
+	return rv.status;
+}
+
+static inline void tiocx_intr_free(nasid_t nasid, int widget,
+				   struct sn_irq_info *sn_irq_info)
+{
+	struct ia64_sal_retval rv;
+	rv.status = 0;
+	rv.v0 = 0;
+
+	ia64_sal_oemcall_nolock(&rv, SN_SAL_IOIF_INTERRUPT,
+				SAL_INTR_FREE, nasid,
+				widget, sn_irq_info->irq_irq,
+				sn_irq_info->irq_cookie, 0, 0);
+}
+
+struct sn_irq_info *tiocx_irq_alloc(nasid_t nasid, int widget, int irq,
+				    nasid_t req_nasid, int slice)
+{
+	struct sn_irq_info *sn_irq_info;
+	int status;
+	int sn_irq_size = sizeof(struct sn_irq_info);
+
+	if ((nasid & 1) == 0)
+		return NULL;
+
+	sn_irq_info = kzalloc(sn_irq_size, GFP_KERNEL);
+	if (sn_irq_info == NULL)
+		return NULL;
+
+	status = tiocx_intr_alloc(nasid, widget, __pa(sn_irq_info), irq,
+				  req_nasid, slice);
+	if (status) {
+		kfree(sn_irq_info);
+		return NULL;
+	} else {
+		return sn_irq_info;
+	}
+}
+
+void tiocx_irq_free(struct sn_irq_info *sn_irq_info)
+{
+	u64 bridge = (u64) sn_irq_info->irq_bridge;
+	nasid_t nasid = NASID_GET(bridge);
+	int widget;
+
+	if (nasid & 1) {
+		widget = TIO_SWIN_WIDGETNUM(bridge);
+		tiocx_intr_free(nasid, widget, sn_irq_info);
+		kfree(sn_irq_info);
+	}
+}
+
+u64 tiocx_dma_addr(u64 addr)
+{
+	return PHYS_TO_TIODMA(addr);
+}
+
+u64 tiocx_swin_base(int nasid)
+{
+	return TIO_SWIN_BASE(nasid, TIOCX_CORELET);
+}
+
+EXPORT_SYMBOL(cx_driver_register);
+EXPORT_SYMBOL(cx_driver_unregister);
+EXPORT_SYMBOL(cx_device_register);
+EXPORT_SYMBOL(cx_device_unregister);
+EXPORT_SYMBOL(tiocx_irq_alloc);
+EXPORT_SYMBOL(tiocx_irq_free);
+EXPORT_SYMBOL(tiocx_bus_type);
+EXPORT_SYMBOL(tiocx_dma_addr);
+EXPORT_SYMBOL(tiocx_swin_base);
+
+static void tio_conveyor_set(nasid_t nasid, int enable_flag)
+{
+	u64 ice_frz;
+	u64 disable_cb = (1ull << 61);
+
+	if (!(nasid & 1))
+		return;
+
+	ice_frz = REMOTE_HUB_L(nasid, TIO_ICE_FRZ_CFG);
+	if (enable_flag) {
+		if (!(ice_frz & disable_cb))	/* already enabled */
+			return;
+		ice_frz &= ~disable_cb;
+	} else {
+		if (ice_frz & disable_cb)	/* already disabled */
+			return;
+		ice_frz |= disable_cb;
+	}
+	DBG(KERN_ALERT "TIO_ICE_FRZ_CFG= 0x%lx\n", ice_frz);
+	REMOTE_HUB_S(nasid, TIO_ICE_FRZ_CFG, ice_frz);
+}
+
+#define tio_conveyor_enable(nasid) tio_conveyor_set(nasid, 1)
+#define tio_conveyor_disable(nasid) tio_conveyor_set(nasid, 0)
+
+static void tio_corelet_reset(nasid_t nasid, int corelet)
+{
+	if (!(nasid & 1))
+		return;
+
+	REMOTE_HUB_S(nasid, TIO_ICE_PMI_TX_CFG, 1 << corelet);
+	udelay(2000);
+	REMOTE_HUB_S(nasid, TIO_ICE_PMI_TX_CFG, 0);
+	udelay(2000);
+}
+
+static int is_fpga_tio(int nasid, int *bt)
+{
+	u16 uninitialized_var(ioboard_type);	/* GCC be quiet */
+	long rc;
+
+	rc = ia64_sn_sysctl_ioboard_get(nasid, &ioboard_type);
+	if (rc) {
+		printk(KERN_WARNING "ia64_sn_sysctl_ioboard_get failed: %ld\n",
+		       rc);
+		return 0;
+	}
+
+	switch (ioboard_type) {
+	case L1_BRICKTYPE_SA:
+	case L1_BRICKTYPE_ATHENA:
+	case L1_BOARDTYPE_DAYTONA:
+		*bt = ioboard_type;
+		return 1;
+	}
+
+	return 0;
+}
+
+static int bitstream_loaded(nasid_t nasid)
+{
+	u64 cx_credits;
+
+	cx_credits = REMOTE_HUB_L(nasid, TIO_ICE_PMI_TX_DYN_CREDIT_STAT_CB3);
+	cx_credits &= TIO_ICE_PMI_TX_DYN_CREDIT_STAT_CB3_CREDIT_CNT_MASK;
+	DBG("cx_credits= 0x%lx\n", cx_credits);
+
+	return (cx_credits == 0xf) ? 1 : 0;
+}
+
+static int tiocx_reload(struct cx_dev *cx_dev)
+{
+	int part_num = CX_DEV_NONE;
+	int mfg_num = CX_DEV_NONE;
+	nasid_t nasid = cx_dev->cx_id.nasid;
+
+	if (bitstream_loaded(nasid)) {
+		u64 cx_id;
+		int rv;
+
+		rv = ia64_sn_sysctl_tio_clock_reset(nasid);
+		if (rv) {
+			printk(KERN_ALERT "CX port JTAG reset failed.\n");
+		} else {
+			cx_id = *(volatile u64 *)
+				(TIO_SWIN_BASE(nasid, TIOCX_CORELET) +
+					  WIDGET_ID);
+			part_num = XWIDGET_PART_NUM(cx_id);
+			mfg_num = XWIDGET_MFG_NUM(cx_id);
+			DBG("part= 0x%x, mfg= 0x%x\n", part_num, mfg_num);
+			/* just ignore it if it's a CE */
+			if (part_num == TIO_CE_ASIC_PARTNUM)
+				return 0;
+		}
+	}
+
+	cx_dev->cx_id.part_num = part_num;
+	cx_dev->cx_id.mfg_num = mfg_num;
+
+	/*
+	 * Delete old device and register the new one.  It's ok if
+	 * part_num/mfg_num == CX_DEV_NONE.  We want to register
+	 * devices in the table even if a bitstream isn't loaded.
+	 * That allows use to see that a bitstream isn't loaded via
+	 * TIOCX_IOCTL_DEV_LIST.
+	 */
+	return cx_device_reload(cx_dev);
+}
+
+static ssize_t show_cxdev_control(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+
+	return sprintf(buf, "0x%x 0x%x 0x%x 0x%x\n",
+		       cx_dev->cx_id.nasid,
+		       cx_dev->cx_id.part_num, cx_dev->cx_id.mfg_num,
+		       cx_dev->bt);
+}
+
+static ssize_t store_cxdev_control(struct device *dev, struct device_attribute *attr, const char *buf,
+				   size_t count)
+{
+	int n;
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+
+	if (!capable(CAP_SYS_ADMIN))
+		return -EPERM;
+
+	if (count <= 0)
+		return 0;
+
+	n = simple_strtoul(buf, NULL, 0);
+
+	switch (n) {
+	case 1:
+		tio_corelet_reset(cx_dev->cx_id.nasid, TIOCX_CORELET);
+		tiocx_reload(cx_dev);
+		break;
+	case 2:
+		tiocx_reload(cx_dev);
+		break;
+	case 3:
+		tio_corelet_reset(cx_dev->cx_id.nasid, TIOCX_CORELET);
+		break;
+	default:
+		break;
+	}
+
+	return count;
+}
+
+DEVICE_ATTR(cxdev_control, 0644, show_cxdev_control, store_cxdev_control);
+
+static int __init tiocx_init(void)
+{
+	cnodeid_t cnodeid;
+	int found_tiocx_device = 0;
+	int err;
+
+	if (!ia64_platform_is("sn2"))
+		return 0;
+
+	err = bus_register(&tiocx_bus_type);
+	if (err)
+		return err;
+
+	for (cnodeid = 0; cnodeid < num_cnodes; cnodeid++) {
+		nasid_t nasid;
+		int bt;
+
+		nasid = cnodeid_to_nasid(cnodeid);
+
+		if ((nasid & 0x1) && is_fpga_tio(nasid, &bt)) {
+			struct hubdev_info *hubdev;
+			struct xwidget_info *widgetp;
+
+			DBG("Found TIO at nasid 0x%x\n", nasid);
+
+			hubdev =
+			    (struct hubdev_info *)(NODEPDA(cnodeid)->pdinfo);
+
+			widgetp = &hubdev->hdi_xwidget_info[TIOCX_CORELET];
+
+			/* The CE hangs off of the CX port but is not an FPGA */
+			if (widgetp->xwi_hwid.part_num == TIO_CE_ASIC_PARTNUM)
+				continue;
+
+			tio_corelet_reset(nasid, TIOCX_CORELET);
+			tio_conveyor_enable(nasid);
+
+			if (cx_device_register
+			    (nasid, widgetp->xwi_hwid.part_num,
+			     widgetp->xwi_hwid.mfg_num, hubdev, bt) < 0)
+				return -ENXIO;
+			else
+				found_tiocx_device++;
+		}
+	}
+
+	/* It's ok if we find zero devices. */
+	DBG("found_tiocx_device= %d\n", found_tiocx_device);
+
+	return 0;
+}
+
+static int cx_remove_device(struct device * dev, void * data)
+{
+	struct cx_dev *cx_dev = to_cx_dev(dev);
+	device_remove_file(dev, &dev_attr_cxdev_control);
+	cx_device_unregister(cx_dev);
+	return 0;
+}
+
+static void __exit tiocx_exit(void)
+{
+	DBG("tiocx_exit\n");
+
+	/*
+	 * Unregister devices.
+	 */
+	bus_for_each_dev(&tiocx_bus_type, NULL, NULL, cx_remove_device);
+	bus_unregister(&tiocx_bus_type);
+}
+
+fs_initcall(tiocx_init);
+module_exit(tiocx_exit);
+
+/************************************************************************
+ * Module licensing and description
+ ************************************************************************/
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Bruce Losure <blosure@sgi.com>");
+MODULE_DESCRIPTION("TIOCX module");
+MODULE_SUPPORTED_DEVICE(DEVICE_NAME);
diff --git a/kernel/arch/ia64/sn/pci/Makefile b/kernel/arch/ia64/sn/pci/Makefile
new file mode 100644
index 000000000..df2a90145
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/Makefile
@@ -0,0 +1,12 @@
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# Copyright (C) 2000-2004 Silicon Graphics, Inc.  All Rights Reserved.
+#
+# Makefile for the sn pci general routines.
+
+ccflags-y := -Iarch/ia64/sn/include
+
+obj-y := pci_dma.o tioca_provider.o tioce_provider.o pcibr/
diff --git a/kernel/arch/ia64/sn/pci/pci_dma.c b/kernel/arch/ia64/sn/pci/pci_dma.c
new file mode 100644
index 000000000..d0853e8e8
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pci_dma.c
@@ -0,0 +1,487 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2000,2002-2005 Silicon Graphics, Inc. All rights reserved.
+ *
+ * Routines for PCI DMA mapping.  See Documentation/DMA-API.txt for
+ * a description of how these routines should be used.
+ */
+
+#include <linux/gfp.h>
+#include <linux/module.h>
+#include <linux/dma-mapping.h>
+#include <asm/dma.h>
+#include <asm/sn/intr.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/sn_sal.h>
+
+#define SG_ENT_VIRT_ADDRESS(sg)	(sg_virt((sg)))
+#define SG_ENT_PHYS_ADDRESS(SG)	virt_to_phys(SG_ENT_VIRT_ADDRESS(SG))
+
+/**
+ * sn_dma_supported - test a DMA mask
+ * @dev: device to test
+ * @mask: DMA mask to test
+ *
+ * Return whether the given PCI device DMA address mask can be supported
+ * properly.  For example, if your device can only drive the low 24-bits
+ * during PCI bus mastering, then you would pass 0x00ffffff as the mask to
+ * this function.  Of course, SN only supports devices that have 32 or more
+ * address bits when using the PMU.
+ */
+static int sn_dma_supported(struct device *dev, u64 mask)
+{
+	BUG_ON(!dev_is_pci(dev));
+
+	if (mask < 0x7fffffff)
+		return 0;
+	return 1;
+}
+
+/**
+ * sn_dma_set_mask - set the DMA mask
+ * @dev: device to set
+ * @dma_mask: new mask
+ *
+ * Set @dev's DMA mask if the hw supports it.
+ */
+int sn_dma_set_mask(struct device *dev, u64 dma_mask)
+{
+	BUG_ON(!dev_is_pci(dev));
+
+	if (!sn_dma_supported(dev, dma_mask))
+		return 0;
+
+	*dev->dma_mask = dma_mask;
+	return 1;
+}
+EXPORT_SYMBOL(sn_dma_set_mask);
+
+/**
+ * sn_dma_alloc_coherent - allocate memory for coherent DMA
+ * @dev: device to allocate for
+ * @size: size of the region
+ * @dma_handle: DMA (bus) address
+ * @flags: memory allocation flags
+ *
+ * dma_alloc_coherent() returns a pointer to a memory region suitable for
+ * coherent DMA traffic to/from a PCI device.  On SN platforms, this means
+ * that @dma_handle will have the %PCIIO_DMA_CMD flag set.
+ *
+ * This interface is usually used for "command" streams (e.g. the command
+ * queue for a SCSI controller).  See Documentation/DMA-API.txt for
+ * more information.
+ */
+static void *sn_dma_alloc_coherent(struct device *dev, size_t size,
+				   dma_addr_t * dma_handle, gfp_t flags,
+				   struct dma_attrs *attrs)
+{
+	void *cpuaddr;
+	unsigned long phys_addr;
+	int node;
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+
+	BUG_ON(!dev_is_pci(dev));
+
+	/*
+	 * Allocate the memory.
+	 */
+	node = pcibus_to_node(pdev->bus);
+	if (likely(node >=0)) {
+		struct page *p = alloc_pages_exact_node(node,
+						flags, get_order(size));
+
+		if (likely(p))
+			cpuaddr = page_address(p);
+		else
+			return NULL;
+	} else
+		cpuaddr = (void *)__get_free_pages(flags, get_order(size));
+
+	if (unlikely(!cpuaddr))
+		return NULL;
+
+	memset(cpuaddr, 0x0, size);
+
+	/* physical addr. of the memory we just got */
+	phys_addr = __pa(cpuaddr);
+
+	/*
+	 * 64 bit address translations should never fail.
+	 * 32 bit translations can fail if there are insufficient mapping
+	 * resources.
+	 */
+
+	*dma_handle = provider->dma_map_consistent(pdev, phys_addr, size,
+						   SN_DMA_ADDR_PHYS);
+	if (!*dma_handle) {
+		printk(KERN_ERR "%s: out of ATEs\n", __func__);
+		free_pages((unsigned long)cpuaddr, get_order(size));
+		return NULL;
+	}
+
+	return cpuaddr;
+}
+
+/**
+ * sn_pci_free_coherent - free memory associated with coherent DMAable region
+ * @dev: device to free for
+ * @size: size to free
+ * @cpu_addr: kernel virtual address to free
+ * @dma_handle: DMA address associated with this region
+ *
+ * Frees the memory allocated by dma_alloc_coherent(), potentially unmapping
+ * any associated IOMMU mappings.
+ */
+static void sn_dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
+				 dma_addr_t dma_handle, struct dma_attrs *attrs)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+
+	BUG_ON(!dev_is_pci(dev));
+
+	provider->dma_unmap(pdev, dma_handle, 0);
+	free_pages((unsigned long)cpu_addr, get_order(size));
+}
+
+/**
+ * sn_dma_map_single_attrs - map a single page for DMA
+ * @dev: device to map for
+ * @cpu_addr: kernel virtual address of the region to map
+ * @size: size of the region
+ * @direction: DMA direction
+ * @attrs: optional dma attributes
+ *
+ * Map the region pointed to by @cpu_addr for DMA and return the
+ * DMA address.
+ *
+ * We map this to the one step pcibr_dmamap_trans interface rather than
+ * the two step pcibr_dmamap_alloc/pcibr_dmamap_addr because we have
+ * no way of saving the dmamap handle from the alloc to later free
+ * (which is pretty much unacceptable).
+ *
+ * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
+ * dma_map_consistent() so that writes force a flush of pending DMA.
+ * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
+ * Document Number: 007-4763-001)
+ *
+ * TODO: simplify our interface;
+ *       figure out how to save dmamap handle so can use two step.
+ */
+static dma_addr_t sn_dma_map_page(struct device *dev, struct page *page,
+				  unsigned long offset, size_t size,
+				  enum dma_data_direction dir,
+				  struct dma_attrs *attrs)
+{
+	void *cpu_addr = page_address(page) + offset;
+	dma_addr_t dma_addr;
+	unsigned long phys_addr;
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+	int dmabarr;
+
+	dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
+
+	BUG_ON(!dev_is_pci(dev));
+
+	phys_addr = __pa(cpu_addr);
+	if (dmabarr)
+		dma_addr = provider->dma_map_consistent(pdev, phys_addr,
+							size, SN_DMA_ADDR_PHYS);
+	else
+		dma_addr = provider->dma_map(pdev, phys_addr, size,
+					     SN_DMA_ADDR_PHYS);
+
+	if (!dma_addr) {
+		printk(KERN_ERR "%s: out of ATEs\n", __func__);
+		return 0;
+	}
+	return dma_addr;
+}
+
+/**
+ * sn_dma_unmap_single_attrs - unamp a DMA mapped page
+ * @dev: device to sync
+ * @dma_addr: DMA address to sync
+ * @size: size of region
+ * @direction: DMA direction
+ * @attrs: optional dma attributes
+ *
+ * This routine is supposed to sync the DMA region specified
+ * by @dma_handle into the coherence domain.  On SN, we're always cache
+ * coherent, so we just need to free any ATEs associated with this mapping.
+ */
+static void sn_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
+			      size_t size, enum dma_data_direction dir,
+			      struct dma_attrs *attrs)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+
+	BUG_ON(!dev_is_pci(dev));
+
+	provider->dma_unmap(pdev, dma_addr, dir);
+}
+
+/**
+ * sn_dma_unmap_sg - unmap a DMA scatterlist
+ * @dev: device to unmap
+ * @sg: scatterlist to unmap
+ * @nhwentries: number of scatterlist entries
+ * @direction: DMA direction
+ * @attrs: optional dma attributes
+ *
+ * Unmap a set of streaming mode DMA translations.
+ */
+static void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
+			    int nhwentries, enum dma_data_direction dir,
+			    struct dma_attrs *attrs)
+{
+	int i;
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+	struct scatterlist *sg;
+
+	BUG_ON(!dev_is_pci(dev));
+
+	for_each_sg(sgl, sg, nhwentries, i) {
+		provider->dma_unmap(pdev, sg->dma_address, dir);
+		sg->dma_address = (dma_addr_t) NULL;
+		sg->dma_length = 0;
+	}
+}
+
+/**
+ * sn_dma_map_sg - map a scatterlist for DMA
+ * @dev: device to map for
+ * @sg: scatterlist to map
+ * @nhwentries: number of entries
+ * @direction: direction of the DMA transaction
+ * @attrs: optional dma attributes
+ *
+ * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
+ * dma_map_consistent() so that writes force a flush of pending DMA.
+ * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
+ * Document Number: 007-4763-001)
+ *
+ * Maps each entry of @sg for DMA.
+ */
+static int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl,
+			 int nhwentries, enum dma_data_direction dir,
+			 struct dma_attrs *attrs)
+{
+	unsigned long phys_addr;
+	struct scatterlist *saved_sg = sgl, *sg;
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
+	int i;
+	int dmabarr;
+
+	dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
+
+	BUG_ON(!dev_is_pci(dev));
+
+	/*
+	 * Setup a DMA address for each entry in the scatterlist.
+	 */
+	for_each_sg(sgl, sg, nhwentries, i) {
+		dma_addr_t dma_addr;
+		phys_addr = SG_ENT_PHYS_ADDRESS(sg);
+		if (dmabarr)
+			dma_addr = provider->dma_map_consistent(pdev,
+								phys_addr,
+								sg->length,
+								SN_DMA_ADDR_PHYS);
+		else
+			dma_addr = provider->dma_map(pdev, phys_addr,
+						     sg->length,
+						     SN_DMA_ADDR_PHYS);
+
+		sg->dma_address = dma_addr;
+		if (!sg->dma_address) {
+			printk(KERN_ERR "%s: out of ATEs\n", __func__);
+
+			/*
+			 * Free any successfully allocated entries.
+			 */
+			if (i > 0)
+				sn_dma_unmap_sg(dev, saved_sg, i, dir, attrs);
+			return 0;
+		}
+
+		sg->dma_length = sg->length;
+	}
+
+	return nhwentries;
+}
+
+static void sn_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
+				       size_t size, enum dma_data_direction dir)
+{
+	BUG_ON(!dev_is_pci(dev));
+}
+
+static void sn_dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle,
+					  size_t size,
+					  enum dma_data_direction dir)
+{
+	BUG_ON(!dev_is_pci(dev));
+}
+
+static void sn_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
+				   int nelems, enum dma_data_direction dir)
+{
+	BUG_ON(!dev_is_pci(dev));
+}
+
+static void sn_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
+				      int nelems, enum dma_data_direction dir)
+{
+	BUG_ON(!dev_is_pci(dev));
+}
+
+static int sn_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
+{
+	return 0;
+}
+
+u64 sn_dma_get_required_mask(struct device *dev)
+{
+	return DMA_BIT_MASK(64);
+}
+EXPORT_SYMBOL_GPL(sn_dma_get_required_mask);
+
+char *sn_pci_get_legacy_mem(struct pci_bus *bus)
+{
+	if (!SN_PCIBUS_BUSSOFT(bus))
+		return ERR_PTR(-ENODEV);
+
+	return (char *)(SN_PCIBUS_BUSSOFT(bus)->bs_legacy_mem | __IA64_UNCACHED_OFFSET);
+}
+
+int sn_pci_legacy_read(struct pci_bus *bus, u16 port, u32 *val, u8 size)
+{
+	unsigned long addr;
+	int ret;
+	struct ia64_sal_retval isrv;
+
+	/*
+	 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
+	 * around hw issues at the pci bus level.  SGI proms older than
+	 * 4.10 don't implement this.
+	 */
+
+	SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
+		 pci_domain_nr(bus), bus->number,
+		 0, /* io */
+		 0, /* read */
+		 port, size, __pa(val));
+
+	if (isrv.status == 0)
+		return size;
+
+	/*
+	 * If the above failed, retry using the SAL_PROBE call which should
+	 * be present in all proms (but which cannot work round PCI chipset
+	 * bugs).  This code is retained for compatibility with old
+	 * pre-4.10 proms, and should be removed at some point in the future.
+	 */
+
+	if (!SN_PCIBUS_BUSSOFT(bus))
+		return -ENODEV;
+
+	addr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
+	addr += port;
+
+	ret = ia64_sn_probe_mem(addr, (long)size, (void *)val);
+
+	if (ret == 2)
+		return -EINVAL;
+
+	if (ret == 1)
+		*val = -1;
+
+	return size;
+}
+
+int sn_pci_legacy_write(struct pci_bus *bus, u16 port, u32 val, u8 size)
+{
+	int ret = size;
+	unsigned long paddr;
+	unsigned long *addr;
+	struct ia64_sal_retval isrv;
+
+	/*
+	 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
+	 * around hw issues at the pci bus level.  SGI proms older than
+	 * 4.10 don't implement this.
+	 */
+
+	SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
+		 pci_domain_nr(bus), bus->number,
+		 0, /* io */
+		 1, /* write */
+		 port, size, __pa(&val));
+
+	if (isrv.status == 0)
+		return size;
+
+	/*
+	 * If the above failed, retry using the SAL_PROBE call which should
+	 * be present in all proms (but which cannot work round PCI chipset
+	 * bugs).  This code is retained for compatibility with old
+	 * pre-4.10 proms, and should be removed at some point in the future.
+	 */
+
+	if (!SN_PCIBUS_BUSSOFT(bus)) {
+		ret = -ENODEV;
+		goto out;
+	}
+
+	/* Put the phys addr in uncached space */
+	paddr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
+	paddr += port;
+	addr = (unsigned long *)paddr;
+
+	switch (size) {
+	case 1:
+		*(volatile u8 *)(addr) = (u8)(val);
+		break;
+	case 2:
+		*(volatile u16 *)(addr) = (u16)(val);
+		break;
+	case 4:
+		*(volatile u32 *)(addr) = (u32)(val);
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+ out:
+	return ret;
+}
+
+static struct dma_map_ops sn_dma_ops = {
+	.alloc			= sn_dma_alloc_coherent,
+	.free			= sn_dma_free_coherent,
+	.map_page		= sn_dma_map_page,
+	.unmap_page		= sn_dma_unmap_page,
+	.map_sg			= sn_dma_map_sg,
+	.unmap_sg		= sn_dma_unmap_sg,
+	.sync_single_for_cpu 	= sn_dma_sync_single_for_cpu,
+	.sync_sg_for_cpu	= sn_dma_sync_sg_for_cpu,
+	.sync_single_for_device = sn_dma_sync_single_for_device,
+	.sync_sg_for_device	= sn_dma_sync_sg_for_device,
+	.mapping_error		= sn_dma_mapping_error,
+	.dma_supported		= sn_dma_supported,
+};
+
+void sn_dma_init(void)
+{
+	dma_ops = &sn_dma_ops;
+}
diff --git a/kernel/arch/ia64/sn/pci/pcibr/Makefile b/kernel/arch/ia64/sn/pci/pcibr/Makefile
new file mode 100644
index 000000000..396bcae36
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pcibr/Makefile
@@ -0,0 +1,13 @@
+#
+# This file is subject to the terms and conditions of the GNU General Public
+# License.  See the file "COPYING" in the main directory of this archive
+# for more details.
+#
+# Copyright (C) 2002-2004 Silicon Graphics, Inc.  All Rights Reserved.
+#
+# Makefile for the sn2 io routines.
+
+ccflags-y := -Iarch/ia64/sn/include
+
+obj-y				+=  pcibr_dma.o pcibr_reg.o \
+				    pcibr_ate.o pcibr_provider.o
diff --git a/kernel/arch/ia64/sn/pci/pcibr/pcibr_ate.c b/kernel/arch/ia64/sn/pci/pcibr/pcibr_ate.c
new file mode 100644
index 000000000..5bc34eac9
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pcibr/pcibr_ate.c
@@ -0,0 +1,177 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/types.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+
+int pcibr_invalidate_ate;	/* by default don't invalidate ATE on free */
+
+/*
+ * mark_ate: Mark the ate as either free or inuse.
+ */
+static void mark_ate(struct ate_resource *ate_resource, int start, int number,
+		     u64 value)
+{
+	u64 *ate = ate_resource->ate;
+	int index;
+	int length = 0;
+
+	for (index = start; length < number; index++, length++)
+		ate[index] = value;
+}
+
+/*
+ * find_free_ate:  Find the first free ate index starting from the given
+ *		   index for the desired consecutive count.
+ */
+static int find_free_ate(struct ate_resource *ate_resource, int start,
+			 int count)
+{
+	u64 *ate = ate_resource->ate;
+	int index;
+	int start_free;
+
+	for (index = start; index < ate_resource->num_ate;) {
+		if (!ate[index]) {
+			int i;
+			int free;
+			free = 0;
+			start_free = index;	/* Found start free ate */
+			for (i = start_free; i < ate_resource->num_ate; i++) {
+				if (!ate[i]) {	/* This is free */
+					if (++free == count)
+						return start_free;
+				} else {
+					index = i + 1;
+					break;
+				}
+			}
+			if (i >= ate_resource->num_ate)
+				return -1;
+		} else
+			index++;	/* Try next ate */
+	}
+
+	return -1;
+}
+
+/*
+ * free_ate_resource:  Free the requested number of ATEs.
+ */
+static inline void free_ate_resource(struct ate_resource *ate_resource,
+				     int start)
+{
+	mark_ate(ate_resource, start, ate_resource->ate[start], 0);
+	if ((ate_resource->lowest_free_index > start) ||
+	    (ate_resource->lowest_free_index < 0))
+		ate_resource->lowest_free_index = start;
+}
+
+/*
+ * alloc_ate_resource:  Allocate the requested number of ATEs.
+ */
+static inline int alloc_ate_resource(struct ate_resource *ate_resource,
+				     int ate_needed)
+{
+	int start_index;
+
+	/*
+	 * Check for ate exhaustion.
+	 */
+	if (ate_resource->lowest_free_index < 0)
+		return -1;
+
+	/*
+	 * Find the required number of free consecutive ates.
+	 */
+	start_index =
+	    find_free_ate(ate_resource, ate_resource->lowest_free_index,
+			  ate_needed);
+	if (start_index >= 0)
+		mark_ate(ate_resource, start_index, ate_needed, ate_needed);
+
+	ate_resource->lowest_free_index =
+	    find_free_ate(ate_resource, ate_resource->lowest_free_index, 1);
+
+	return start_index;
+}
+
+/*
+ * Allocate "count" contiguous Bridge Address Translation Entries
+ * on the specified bridge to be used for PCI to XTALK mappings.
+ * Indices in rm map range from 1..num_entries.  Indices returned
+ * to caller range from 0..num_entries-1.
+ *
+ * Return the start index on success, -1 on failure.
+ */
+int pcibr_ate_alloc(struct pcibus_info *pcibus_info, int count)
+{
+	int status;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pcibus_info->pbi_lock, flags);
+	status = alloc_ate_resource(&pcibus_info->pbi_int_ate_resource, count);
+	spin_unlock_irqrestore(&pcibus_info->pbi_lock, flags);
+
+	return status;
+}
+
+/*
+ * Setup an Address Translation Entry as specified.  Use either the Bridge
+ * internal maps or the external map RAM, as appropriate.
+ */
+static inline u64 __iomem *pcibr_ate_addr(struct pcibus_info *pcibus_info,
+				       int ate_index)
+{
+	if (ate_index < pcibus_info->pbi_int_ate_size) {
+		return pcireg_int_ate_addr(pcibus_info, ate_index);
+	}
+	panic("pcibr_ate_addr: invalid ate_index 0x%x", ate_index);
+}
+
+/*
+ * Update the ate.
+ */
+void inline
+ate_write(struct pcibus_info *pcibus_info, int ate_index, int count,
+	  volatile u64 ate)
+{
+	while (count-- > 0) {
+		if (ate_index < pcibus_info->pbi_int_ate_size) {
+			pcireg_int_ate_set(pcibus_info, ate_index, ate);
+		} else {
+			panic("ate_write: invalid ate_index 0x%x", ate_index);
+		}
+		ate_index++;
+		ate += IOPGSIZE;
+	}
+
+	pcireg_tflush_get(pcibus_info);	/* wait until Bridge PIO complete */
+}
+
+void pcibr_ate_free(struct pcibus_info *pcibus_info, int index)
+{
+
+	volatile u64 ate;
+	int count;
+	unsigned long flags;
+
+	if (pcibr_invalidate_ate) {
+		/* For debugging purposes, clear the valid bit in the ATE */
+		ate = *pcibr_ate_addr(pcibus_info, index);
+		count = pcibus_info->pbi_int_ate_resource.ate[index];
+		ate_write(pcibus_info, index, count, (ate & ~PCI32_ATE_V));
+	}
+
+	spin_lock_irqsave(&pcibus_info->pbi_lock, flags);
+	free_ate_resource(&pcibus_info->pbi_int_ate_resource, index);
+	spin_unlock_irqrestore(&pcibus_info->pbi_lock, flags);
+}
diff --git a/kernel/arch/ia64/sn/pci/pcibr/pcibr_dma.c b/kernel/arch/ia64/sn/pci/pcibr/pcibr_dma.c
new file mode 100644
index 000000000..1e863b277
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pcibr/pcibr_dma.c
@@ -0,0 +1,413 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2001-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/pci.h>
+#include <linux/export.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/geo.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/pic.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/tiocp.h>
+#include "tio.h"
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+
+extern int sn_ioif_inited;
+
+/* =====================================================================
+ *    DMA MANAGEMENT
+ *
+ *      The Bridge ASIC provides three methods of doing DMA: via a "direct map"
+ *      register available in 32-bit PCI space (which selects a contiguous 2G
+ *	address space on some other widget), via "direct" addressing via 64-bit
+ *      PCI space (all destination information comes from the PCI address,
+ *      including transfer attributes), and via a "mapped" region that allows 
+ *      a bunch of different small mappings to be established with the PMU.
+ *
+ *      For efficiency, we most prefer to use the 32bit direct mapping facility,
+ *      since it requires no resource allocations. The advantage of using the
+ *      PMU over the 64-bit direct is that single-cycle PCI addressing can be
+ *      used; the advantage of using 64-bit direct over PMU addressing is that
+ *      we do not have to allocate entries in the PMU.
+ */
+
+static dma_addr_t
+pcibr_dmamap_ate32(struct pcidev_info *info,
+		   u64 paddr, size_t req_size, u64 flags, int dma_flags)
+{
+
+	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
+	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
+	    pdi_pcibus_info;
+	u8 internal_device = (PCI_SLOT(pcidev_info->pdi_host_pcidev_info->
+					    pdi_linux_pcidev->devfn)) - 1;
+	int ate_count;
+	int ate_index;
+	u64 ate_flags = flags | PCI32_ATE_V;
+	u64 ate;
+	u64 pci_addr;
+	u64 xio_addr;
+	u64 offset;
+
+	/* PIC in PCI-X mode does not supports 32bit PageMap mode */
+	if (IS_PIC_SOFT(pcibus_info) && IS_PCIX(pcibus_info)) {
+		return 0;
+	}
+
+	/* Calculate the number of ATEs needed. */
+	if (!(MINIMAL_ATE_FLAG(paddr, req_size))) {
+		ate_count = IOPG((IOPGSIZE - 1)	/* worst case start offset */
+				 +req_size	/* max mapping bytes */
+				 - 1) + 1;	/* round UP */
+	} else {		/* assume requested target is page aligned */
+		ate_count = IOPG(req_size	/* max mapping bytes */
+				 - 1) + 1;	/* round UP */
+	}
+
+	/* Get the number of ATEs required. */
+	ate_index = pcibr_ate_alloc(pcibus_info, ate_count);
+	if (ate_index < 0)
+		return 0;
+
+	/* In PCI-X mode, Prefetch not supported */
+	if (IS_PCIX(pcibus_info))
+		ate_flags &= ~(PCI32_ATE_PREF);
+
+	if (SN_DMA_ADDRTYPE(dma_flags == SN_DMA_ADDR_PHYS))
+		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
+	    					      PHYS_TO_TIODMA(paddr);
+	else
+		xio_addr = paddr;
+
+	offset = IOPGOFF(xio_addr);
+	ate = ate_flags | (xio_addr - offset);
+
+	/* If PIC, put the targetid in the ATE */
+	if (IS_PIC_SOFT(pcibus_info)) {
+		ate |= (pcibus_info->pbi_hub_xid << PIC_ATE_TARGETID_SHFT);
+	}
+
+	/*
+	 * If we're mapping for MSI, set the MSI bit in the ATE.  If it's a
+	 * TIOCP based pci bus, we also need to set the PIO bit in the ATE.
+	 */
+	if (dma_flags & SN_DMA_MSI) {
+		ate |= PCI32_ATE_MSI;
+		if (IS_TIOCP_SOFT(pcibus_info))
+			ate |= PCI32_ATE_PIO;
+	}
+
+	ate_write(pcibus_info, ate_index, ate_count, ate);
+
+	/*
+	 * Set up the DMA mapped Address.
+	 */
+	pci_addr = PCI32_MAPPED_BASE + offset + IOPGSIZE * ate_index;
+
+	/*
+	 * If swap was set in device in pcibr_endian_set()
+	 * we need to turn swapping on.
+	 */
+	if (pcibus_info->pbi_devreg[internal_device] & PCIBR_DEV_SWAP_DIR)
+		ATE_SWAP_ON(pci_addr);
+
+
+	return pci_addr;
+}
+
+static dma_addr_t
+pcibr_dmatrans_direct64(struct pcidev_info * info, u64 paddr,
+			u64 dma_attributes, int dma_flags)
+{
+	struct pcibus_info *pcibus_info = (struct pcibus_info *)
+	    ((info->pdi_host_pcidev_info)->pdi_pcibus_info);
+	u64 pci_addr;
+
+	/* Translate to Crosstalk View of Physical Address */
+	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
+		pci_addr = IS_PIC_SOFT(pcibus_info) ?
+				PHYS_TO_DMA(paddr) :
+				PHYS_TO_TIODMA(paddr);
+	else
+		pci_addr = paddr;
+	pci_addr |= dma_attributes;
+
+	/* Handle Bus mode */
+	if (IS_PCIX(pcibus_info))
+		pci_addr &= ~PCI64_ATTR_PREF;
+
+	/* Handle Bridge Chipset differences */
+	if (IS_PIC_SOFT(pcibus_info)) {
+		pci_addr |=
+		    ((u64) pcibus_info->
+		     pbi_hub_xid << PIC_PCI64_ATTR_TARG_SHFT);
+	} else
+		pci_addr |= (dma_flags & SN_DMA_MSI) ?
+				TIOCP_PCI64_CMDTYPE_MSI :
+				TIOCP_PCI64_CMDTYPE_MEM;
+
+	/* If PCI mode, func zero uses VCHAN0, every other func uses VCHAN1 */
+	if (!IS_PCIX(pcibus_info) && PCI_FUNC(info->pdi_linux_pcidev->devfn))
+		pci_addr |= PCI64_ATTR_VIRTUAL;
+
+	return pci_addr;
+}
+
+static dma_addr_t
+pcibr_dmatrans_direct32(struct pcidev_info * info,
+			u64 paddr, size_t req_size, u64 flags, int dma_flags)
+{
+	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
+	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
+	    pdi_pcibus_info;
+	u64 xio_addr;
+
+	u64 xio_base;
+	u64 offset;
+	u64 endoff;
+
+	if (IS_PCIX(pcibus_info)) {
+		return 0;
+	}
+
+	if (dma_flags & SN_DMA_MSI)
+		return 0;
+
+	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
+		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
+	    					      PHYS_TO_TIODMA(paddr);
+	else
+		xio_addr = paddr;
+
+	xio_base = pcibus_info->pbi_dir_xbase;
+	offset = xio_addr - xio_base;
+	endoff = req_size + offset;
+	if ((req_size > (1ULL << 31)) ||	/* Too Big */
+	    (xio_addr < xio_base) ||	/* Out of range for mappings */
+	    (endoff > (1ULL << 31))) {	/* Too Big */
+		return 0;
+	}
+
+	return PCI32_DIRECT_BASE | offset;
+}
+
+/*
+ * Wrapper routine for freeing DMA maps
+ * DMA mappings for Direct 64 and 32 do not have any DMA maps.
+ */
+void
+pcibr_dma_unmap(struct pci_dev *hwdev, dma_addr_t dma_handle, int direction)
+{
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
+	struct pcibus_info *pcibus_info =
+	    (struct pcibus_info *)pcidev_info->pdi_pcibus_info;
+
+	if (IS_PCI32_MAPPED(dma_handle)) {
+		int ate_index;
+
+		ate_index =
+		    IOPG((ATE_SWAP_OFF(dma_handle) - PCI32_MAPPED_BASE));
+		pcibr_ate_free(pcibus_info, ate_index);
+	}
+}
+
+/*
+ * On SN systems there is a race condition between a PIO read response and 
+ * DMA's.  In rare cases, the read response may beat the DMA, causing the
+ * driver to think that data in memory is complete and meaningful.  This code
+ * eliminates that race.  This routine is called by the PIO read routines
+ * after doing the read.  For PIC this routine then forces a fake interrupt
+ * on another line, which is logically associated with the slot that the PIO
+ * is addressed to.  It then spins while watching the memory location that
+ * the interrupt is targeted to.  When the interrupt response arrives, we 
+ * are sure that the DMA has landed in memory and it is safe for the driver
+ * to proceed.	For TIOCP use the Device(x) Write Request Buffer Flush 
+ * Bridge register since it ensures the data has entered the coherence domain,
+ * unlike the PIC Device(x) Write Request Buffer Flush register.
+ */
+
+void sn_dma_flush(u64 addr)
+{
+	nasid_t nasid;
+	int is_tio;
+	int wid_num;
+	int i, j;
+	unsigned long flags;
+	u64 itte;
+	struct hubdev_info *hubinfo;
+	struct sn_flush_device_kernel *p;
+	struct sn_flush_device_common *common;
+	struct sn_flush_nasid_entry *flush_nasid_list;
+
+	if (!sn_ioif_inited)
+		return;
+
+	nasid = NASID_GET(addr);
+	if (-1 == nasid_to_cnodeid(nasid))
+		return;
+
+	hubinfo = (NODEPDA(nasid_to_cnodeid(nasid)))->pdinfo;
+
+	BUG_ON(!hubinfo);
+
+	flush_nasid_list = &hubinfo->hdi_flush_nasid_list;
+	if (flush_nasid_list->widget_p == NULL)
+		return;
+
+	is_tio = (nasid & 1);
+	if (is_tio) {
+		int itte_index;
+
+		if (TIO_HWIN(addr))
+			itte_index = 0;
+		else if (TIO_BWIN_WINDOWNUM(addr))
+			itte_index = TIO_BWIN_WINDOWNUM(addr);
+		else
+			itte_index = -1;
+
+		if (itte_index >= 0) {
+			itte = flush_nasid_list->iio_itte[itte_index];
+			if (! TIO_ITTE_VALID(itte))
+				return;
+			wid_num = TIO_ITTE_WIDGET(itte);
+		} else
+			wid_num = TIO_SWIN_WIDGETNUM(addr);
+	} else {
+		if (BWIN_WINDOWNUM(addr)) {
+			itte = flush_nasid_list->iio_itte[BWIN_WINDOWNUM(addr)];
+			wid_num = IIO_ITTE_WIDGET(itte);
+		} else
+			wid_num = SWIN_WIDGETNUM(addr);
+	}
+	if (flush_nasid_list->widget_p[wid_num] == NULL)
+		return;
+	p = &flush_nasid_list->widget_p[wid_num][0];
+
+	/* find a matching BAR */
+	for (i = 0; i < DEV_PER_WIDGET; i++,p++) {
+		common = p->common;
+		for (j = 0; j < PCI_ROM_RESOURCE; j++) {
+			if (common->sfdl_bar_list[j].start == 0)
+				break;
+			if (addr >= common->sfdl_bar_list[j].start
+			    && addr <= common->sfdl_bar_list[j].end)
+				break;
+		}
+		if (j < PCI_ROM_RESOURCE && common->sfdl_bar_list[j].start != 0)
+			break;
+	}
+
+	/* if no matching BAR, return without doing anything. */
+	if (i == DEV_PER_WIDGET)
+		return;
+
+	/*
+	 * For TIOCP use the Device(x) Write Request Buffer Flush Bridge
+	 * register since it ensures the data has entered the coherence
+	 * domain, unlike PIC.
+	 */
+	if (is_tio) {
+		/*
+	 	 * Note:  devices behind TIOCE should never be matched in the
+		 * above code, and so the following code is PIC/CP centric.
+		 * If CE ever needs the sn_dma_flush mechanism, we will have
+		 * to account for that here and in tioce_bus_fixup().
+	 	 */
+		u32 tio_id = HUB_L(TIO_IOSPACE_ADDR(nasid, TIO_NODE_ID));
+		u32 revnum = XWIDGET_PART_REV_NUM(tio_id);
+
+		/* TIOCP BRINGUP WAR (PV907516): Don't write buffer flush reg */
+		if ((1 << XWIDGET_PART_REV_NUM_REV(revnum)) & PV907516) {
+			return;
+		} else {
+			pcireg_wrb_flush_get(common->sfdl_pcibus_info,
+					     (common->sfdl_slot - 1));
+		}
+	} else {
+		spin_lock_irqsave(&p->sfdl_flush_lock, flags);
+		*common->sfdl_flush_addr = 0;
+
+		/* force an interrupt. */
+		*(volatile u32 *)(common->sfdl_force_int_addr) = 1;
+
+		/* wait for the interrupt to come back. */
+		while (*(common->sfdl_flush_addr) != 0x10f)
+			cpu_relax();
+
+		/* okay, everything is synched up. */
+		spin_unlock_irqrestore(&p->sfdl_flush_lock, flags);
+	}
+	return;
+}
+
+/*
+ * DMA interfaces.  Called from pci_dma.c routines.
+ */
+
+dma_addr_t
+pcibr_dma_map(struct pci_dev * hwdev, unsigned long phys_addr, size_t size, int dma_flags)
+{
+	dma_addr_t dma_handle;
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
+
+	/* SN cannot support DMA addresses smaller than 32 bits. */
+	if (hwdev->dma_mask < 0x7fffffff) {
+		return 0;
+	}
+
+	if (hwdev->dma_mask == ~0UL) {
+		/*
+		 * Handle the most common case: 64 bit cards.  This
+		 * call should always succeed.
+		 */
+
+		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
+						     PCI64_ATTR_PREF, dma_flags);
+	} else {
+		/* Handle 32-63 bit cards via direct mapping */
+		dma_handle = pcibr_dmatrans_direct32(pcidev_info, phys_addr,
+						     size, 0, dma_flags);
+		if (!dma_handle) {
+			/*
+			 * It is a 32 bit card and we cannot do direct mapping,
+			 * so we use an ATE.
+			 */
+
+			dma_handle = pcibr_dmamap_ate32(pcidev_info, phys_addr,
+							size, PCI32_ATE_PREF,
+							dma_flags);
+		}
+	}
+
+	return dma_handle;
+}
+
+dma_addr_t
+pcibr_dma_map_consistent(struct pci_dev * hwdev, unsigned long phys_addr,
+			 size_t size, int dma_flags)
+{
+	dma_addr_t dma_handle;
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
+
+	if (hwdev->dev.coherent_dma_mask == ~0UL) {
+		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
+					    PCI64_ATTR_BAR, dma_flags);
+	} else {
+		dma_handle = (dma_addr_t) pcibr_dmamap_ate32(pcidev_info,
+						    phys_addr, size,
+						    PCI32_ATE_BAR, dma_flags);
+	}
+
+	return dma_handle;
+}
+
+EXPORT_SYMBOL(sn_dma_flush);
diff --git a/kernel/arch/ia64/sn/pci/pcibr/pcibr_provider.c b/kernel/arch/ia64/sn/pci/pcibr/pcibr_provider.c
new file mode 100644
index 000000000..8dbbef4a4
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pcibr/pcibr_provider.c
@@ -0,0 +1,265 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2001-2004, 2006 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/interrupt.h>
+#include <linux/types.h>
+#include <linux/slab.h>
+#include <linux/pci.h>
+#include <linux/export.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/geo.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/pic.h>
+#include <asm/sn/sn2/sn_hwperf.h>
+#include "xtalk/xwidgetdev.h"
+#include "xtalk/hubdev.h"
+
+int
+sal_pcibr_slot_enable(struct pcibus_info *soft, int device, void *resp,
+                      char **ssdt)
+{
+	struct ia64_sal_retval ret_stuff;
+	u64 busnum;
+	u64 segment;
+
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	segment = soft->pbi_buscommon.bs_persist_segment;
+	busnum = soft->pbi_buscommon.bs_persist_busnum;
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_SLOT_ENABLE, segment,
+			busnum, (u64) device, (u64) resp, (u64)ia64_tpa(ssdt),
+			0, 0);
+
+	return (int)ret_stuff.v0;
+}
+
+int
+sal_pcibr_slot_disable(struct pcibus_info *soft, int device, int action,
+		       void *resp)
+{
+	struct ia64_sal_retval ret_stuff;
+	u64 busnum;
+	u64 segment;
+
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	segment = soft->pbi_buscommon.bs_persist_segment;
+	busnum = soft->pbi_buscommon.bs_persist_busnum;
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_SLOT_DISABLE,
+			segment, busnum, (u64) device, (u64) action,
+			(u64) resp, 0, 0);
+
+	return (int)ret_stuff.v0;
+}
+
+static int sal_pcibr_error_interrupt(struct pcibus_info *soft)
+{
+	struct ia64_sal_retval ret_stuff;
+	u64 busnum;
+	int segment;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	segment = soft->pbi_buscommon.bs_persist_segment;
+	busnum = soft->pbi_buscommon.bs_persist_busnum;
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_ERROR_INTERRUPT,
+			(u64) segment, (u64) busnum, 0, 0, 0, 0, 0);
+
+	return (int)ret_stuff.v0;
+}
+
+u16 sn_ioboard_to_pci_bus(struct pci_bus *pci_bus)
+{
+	long rc;
+	u16 uninitialized_var(ioboard);		/* GCC be quiet */
+	nasid_t nasid = NASID_GET(SN_PCIBUS_BUSSOFT(pci_bus)->bs_base);
+
+	rc = ia64_sn_sysctl_ioboard_get(nasid, &ioboard);
+	if (rc) {
+		printk(KERN_WARNING "ia64_sn_sysctl_ioboard_get failed: %ld\n",
+		       rc);
+		return 0;
+	}
+
+	return ioboard;
+}
+
+/* 
+ * PCI Bridge Error interrupt handler.  Gets invoked whenever a PCI 
+ * bridge sends an error interrupt.
+ */
+static irqreturn_t
+pcibr_error_intr_handler(int irq, void *arg)
+{
+	struct pcibus_info *soft = arg;
+
+	if (sal_pcibr_error_interrupt(soft) < 0)
+		panic("pcibr_error_intr_handler(): Fatal Bridge Error");
+
+	return IRQ_HANDLED;
+}
+
+void *
+pcibr_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller)
+{
+	int nasid, cnode, j;
+	struct hubdev_info *hubdev_info;
+	struct pcibus_info *soft;
+	struct sn_flush_device_kernel *sn_flush_device_kernel;
+	struct sn_flush_device_common *common;
+
+	if (! IS_PCI_BRIDGE_ASIC(prom_bussoft->bs_asic_type)) {
+		return NULL;
+	}
+
+	/*
+	 * Allocate kernel bus soft and copy from prom.
+	 */
+
+	soft = kmemdup(prom_bussoft, sizeof(struct pcibus_info), GFP_KERNEL);
+	if (!soft) {
+		return NULL;
+	}
+
+	soft->pbi_buscommon.bs_base = (unsigned long)
+		ioremap(REGION_OFFSET(soft->pbi_buscommon.bs_base),
+			sizeof(struct pic));
+
+	spin_lock_init(&soft->pbi_lock);
+
+	/*
+	 * register the bridge's error interrupt handler
+	 */
+	if (request_irq(SGI_PCIASIC_ERROR, pcibr_error_intr_handler,
+			IRQF_SHARED, "PCIBR error", (void *)(soft))) {
+		printk(KERN_WARNING
+		       "pcibr cannot allocate interrupt for error handler\n");
+	}
+	irq_set_handler(SGI_PCIASIC_ERROR, handle_level_irq);
+	sn_set_err_irq_affinity(SGI_PCIASIC_ERROR);
+
+	/* 
+	 * Update the Bridge with the "kernel" pagesize 
+	 */
+	if (PAGE_SIZE < 16384) {
+		pcireg_control_bit_clr(soft, PCIBR_CTRL_PAGE_SIZE);
+	} else {
+		pcireg_control_bit_set(soft, PCIBR_CTRL_PAGE_SIZE);
+	}
+
+	nasid = NASID_GET(soft->pbi_buscommon.bs_base);
+	cnode = nasid_to_cnodeid(nasid);
+	hubdev_info = (struct hubdev_info *)(NODEPDA(cnode)->pdinfo);
+
+	if (hubdev_info->hdi_flush_nasid_list.widget_p) {
+		sn_flush_device_kernel = hubdev_info->hdi_flush_nasid_list.
+		    widget_p[(int)soft->pbi_buscommon.bs_xid];
+		if (sn_flush_device_kernel) {
+			for (j = 0; j < DEV_PER_WIDGET;
+			     j++, sn_flush_device_kernel++) {
+				common = sn_flush_device_kernel->common;
+				if (common->sfdl_slot == -1)
+					continue;
+				if ((common->sfdl_persistent_segment ==
+				     soft->pbi_buscommon.bs_persist_segment) &&
+				     (common->sfdl_persistent_busnum ==
+				     soft->pbi_buscommon.bs_persist_busnum))
+					common->sfdl_pcibus_info =
+					    soft;
+			}
+		}
+	}
+
+	/* Setup the PMU ATE map */
+	soft->pbi_int_ate_resource.lowest_free_index = 0;
+	soft->pbi_int_ate_resource.ate =
+	    kzalloc(soft->pbi_int_ate_size * sizeof(u64), GFP_KERNEL);
+
+	if (!soft->pbi_int_ate_resource.ate) {
+		kfree(soft);
+		return NULL;
+	}
+
+	return soft;
+}
+
+void pcibr_force_interrupt(struct sn_irq_info *sn_irq_info)
+{
+	struct pcidev_info *pcidev_info;
+	struct pcibus_info *pcibus_info;
+	int bit = sn_irq_info->irq_int_bit;
+
+	if (! sn_irq_info->irq_bridge)
+		return;
+
+	pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	if (pcidev_info) {
+		pcibus_info =
+		    (struct pcibus_info *)pcidev_info->pdi_host_pcidev_info->
+		    pdi_pcibus_info;
+		pcireg_force_intr_set(pcibus_info, bit);
+	}
+}
+
+void pcibr_target_interrupt(struct sn_irq_info *sn_irq_info)
+{
+	struct pcidev_info *pcidev_info;
+	struct pcibus_info *pcibus_info;
+	int bit = sn_irq_info->irq_int_bit;
+	u64 xtalk_addr = sn_irq_info->irq_xtalkaddr;
+
+	pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	if (pcidev_info) {
+		pcibus_info =
+		    (struct pcibus_info *)pcidev_info->pdi_host_pcidev_info->
+		    pdi_pcibus_info;
+
+		/* Disable the device's IRQ   */
+		pcireg_intr_enable_bit_clr(pcibus_info, (1 << bit));
+
+		/* Change the device's IRQ    */
+		pcireg_intr_addr_addr_set(pcibus_info, bit, xtalk_addr);
+
+		/* Re-enable the device's IRQ */
+		pcireg_intr_enable_bit_set(pcibus_info, (1 << bit));
+
+		pcibr_force_interrupt(sn_irq_info);
+	}
+}
+
+/*
+ * Provider entries for PIC/CP
+ */
+
+struct sn_pcibus_provider pcibr_provider = {
+	.dma_map = pcibr_dma_map,
+	.dma_map_consistent = pcibr_dma_map_consistent,
+	.dma_unmap = pcibr_dma_unmap,
+	.bus_fixup = pcibr_bus_fixup,
+	.force_interrupt = pcibr_force_interrupt,
+	.target_interrupt = pcibr_target_interrupt
+};
+
+int
+pcibr_init_provider(void)
+{
+	sn_pci_provider[PCIIO_ASIC_TYPE_PIC] = &pcibr_provider;
+	sn_pci_provider[PCIIO_ASIC_TYPE_TIOCP] = &pcibr_provider;
+
+	return 0;
+}
+
+EXPORT_SYMBOL_GPL(sal_pcibr_slot_enable);
+EXPORT_SYMBOL_GPL(sal_pcibr_slot_disable);
+EXPORT_SYMBOL_GPL(sn_ioboard_to_pci_bus);
diff --git a/kernel/arch/ia64/sn/pci/pcibr/pcibr_reg.c b/kernel/arch/ia64/sn/pci/pcibr/pcibr_reg.c
new file mode 100644
index 000000000..8b8bbd51d
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/pcibr/pcibr_reg.c
@@ -0,0 +1,285 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2004 Silicon Graphics, Inc. All rights reserved.
+ */
+
+#include <linux/interrupt.h>
+#include <linux/types.h>
+#include <asm/sn/io.h>
+#include <asm/sn/pcibr_provider.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/pic.h>
+#include <asm/sn/tiocp.h>
+
+union br_ptr {
+	struct tiocp tio;
+	struct pic pic;
+};
+
+/*
+ * Control Register Access -- Read/Write                            0000_0020
+ */
+void pcireg_control_bit_clr(struct pcibus_info *pcibus_info, u64 bits)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			__sn_clrq_relaxed(&ptr->tio.cp_control, bits);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			__sn_clrq_relaxed(&ptr->pic.p_wid_control, bits);
+			break;
+		default:
+			panic
+			    ("pcireg_control_bit_clr: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+void pcireg_control_bit_set(struct pcibus_info *pcibus_info, u64 bits)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			__sn_setq_relaxed(&ptr->tio.cp_control, bits);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			__sn_setq_relaxed(&ptr->pic.p_wid_control, bits);
+			break;
+		default:
+			panic
+			    ("pcireg_control_bit_set: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+/*
+ * PCI/PCIX Target Flush Register Access -- Read Only		    0000_0050
+ */
+u64 pcireg_tflush_get(struct pcibus_info *pcibus_info)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+	u64 ret = 0;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			ret = __sn_readq_relaxed(&ptr->tio.cp_tflush);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			ret = __sn_readq_relaxed(&ptr->pic.p_wid_tflush);
+			break;
+		default:
+			panic
+			    ("pcireg_tflush_get: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+
+	/* Read of the Target Flush should always return zero */
+	if (ret != 0)
+		panic("pcireg_tflush_get:Target Flush failed\n");
+
+	return ret;
+}
+
+/*
+ * Interrupt Status Register Access -- Read Only		    0000_0100
+ */
+u64 pcireg_intr_status_get(struct pcibus_info * pcibus_info)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+	u64 ret = 0;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			ret = __sn_readq_relaxed(&ptr->tio.cp_int_status);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			ret = __sn_readq_relaxed(&ptr->pic.p_int_status);
+			break;
+		default:
+			panic
+			    ("pcireg_intr_status_get: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+	return ret;
+}
+
+/*
+ * Interrupt Enable Register Access -- Read/Write                   0000_0108
+ */
+void pcireg_intr_enable_bit_clr(struct pcibus_info *pcibus_info, u64 bits)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			__sn_clrq_relaxed(&ptr->tio.cp_int_enable, bits);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			__sn_clrq_relaxed(&ptr->pic.p_int_enable, bits);
+			break;
+		default:
+			panic
+			    ("pcireg_intr_enable_bit_clr: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+void pcireg_intr_enable_bit_set(struct pcibus_info *pcibus_info, u64 bits)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			__sn_setq_relaxed(&ptr->tio.cp_int_enable, bits);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			__sn_setq_relaxed(&ptr->pic.p_int_enable, bits);
+			break;
+		default:
+			panic
+			    ("pcireg_intr_enable_bit_set: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+/*
+ * Intr Host Address Register (int_addr) -- Read/Write  0000_0130 - 0000_0168
+ */
+void pcireg_intr_addr_addr_set(struct pcibus_info *pcibus_info, int int_n,
+			       u64 addr)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			__sn_clrq_relaxed(&ptr->tio.cp_int_addr[int_n],
+			    TIOCP_HOST_INTR_ADDR);
+			__sn_setq_relaxed(&ptr->tio.cp_int_addr[int_n],
+			    (addr & TIOCP_HOST_INTR_ADDR));
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			__sn_clrq_relaxed(&ptr->pic.p_int_addr[int_n],
+			    PIC_HOST_INTR_ADDR);
+			__sn_setq_relaxed(&ptr->pic.p_int_addr[int_n],
+			    (addr & PIC_HOST_INTR_ADDR));
+			break;
+		default:
+			panic
+			    ("pcireg_intr_addr_addr_get: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+/*
+ * Force Interrupt Register Access -- Write Only	0000_01C0 - 0000_01F8
+ */
+void pcireg_force_intr_set(struct pcibus_info *pcibus_info, int int_n)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			writeq(1, &ptr->tio.cp_force_pin[int_n]);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			writeq(1, &ptr->pic.p_force_pin[int_n]);
+			break;
+		default:
+			panic
+			    ("pcireg_force_intr_set: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+/*
+ * Device(x) Write Buffer Flush Reg Access -- Read Only 0000_0240 - 0000_0258
+ */
+u64 pcireg_wrb_flush_get(struct pcibus_info *pcibus_info, int device)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+	u64 ret = 0;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			ret =
+			    __sn_readq_relaxed(&ptr->tio.cp_wr_req_buf[device]);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			ret =
+			    __sn_readq_relaxed(&ptr->pic.p_wr_req_buf[device]);
+			break;
+		default:
+		      panic("pcireg_wrb_flush_get: unknown bridgetype bridge 0x%p", ptr);
+		}
+
+	}
+	/* Read of the Write Buffer Flush should always return zero */
+	return ret;
+}
+
+void pcireg_int_ate_set(struct pcibus_info *pcibus_info, int ate_index,
+			u64 val)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			writeq(val, &ptr->tio.cp_int_ate_ram[ate_index]);
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			writeq(val, &ptr->pic.p_int_ate_ram[ate_index]);
+			break;
+		default:
+			panic
+			    ("pcireg_int_ate_set: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+}
+
+u64 __iomem *pcireg_int_ate_addr(struct pcibus_info *pcibus_info, int ate_index)
+{
+	union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base;
+	u64 __iomem *ret = NULL;
+
+	if (pcibus_info) {
+		switch (pcibus_info->pbi_bridge_type) {
+		case PCIBR_BRIDGETYPE_TIOCP:
+			ret = &ptr->tio.cp_int_ate_ram[ate_index];
+			break;
+		case PCIBR_BRIDGETYPE_PIC:
+			ret = &ptr->pic.p_int_ate_ram[ate_index];
+			break;
+		default:
+			panic
+			    ("pcireg_int_ate_addr: unknown bridgetype bridge 0x%p",
+			     ptr);
+		}
+	}
+	return ret;
+}
diff --git a/kernel/arch/ia64/sn/pci/tioca_provider.c b/kernel/arch/ia64/sn/pci/tioca_provider.c
new file mode 100644
index 000000000..a70b11fd5
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/tioca_provider.c
@@ -0,0 +1,677 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2003-2005 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/bitmap.h>
+#include <linux/slab.h>
+#include <linux/export.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/io.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/tioca_provider.h>
+
+u32 tioca_gart_found;
+EXPORT_SYMBOL(tioca_gart_found);	/* used by agp-sgi */
+
+LIST_HEAD(tioca_list);
+EXPORT_SYMBOL(tioca_list);	/* used by agp-sgi */
+
+static int tioca_gart_init(struct tioca_kernel *);
+
+/**
+ * tioca_gart_init - Initialize SGI TIOCA GART
+ * @tioca_common: ptr to common prom/kernel struct identifying the 
+ *
+ * If the indicated tioca has devices present, initialize its associated
+ * GART MMR's and kernel memory.
+ */
+static int
+tioca_gart_init(struct tioca_kernel *tioca_kern)
+{
+	u64 ap_reg;
+	u64 offset;
+	struct page *tmp;
+	struct tioca_common *tioca_common;
+	struct tioca __iomem *ca_base;
+
+	tioca_common = tioca_kern->ca_common;
+	ca_base = (struct tioca __iomem *)tioca_common->ca_common.bs_base;
+
+	if (list_empty(tioca_kern->ca_devices))
+		return 0;
+
+	ap_reg = 0;
+
+	/*
+	 * Validate aperature size
+	 */
+
+	switch (CA_APERATURE_SIZE >> 20) {
+	case 4:
+		ap_reg |= (0x3ff << CA_GART_AP_SIZE_SHFT);	/* 4MB */
+		break;
+	case 8:
+		ap_reg |= (0x3fe << CA_GART_AP_SIZE_SHFT);	/* 8MB */
+		break;
+	case 16:
+		ap_reg |= (0x3fc << CA_GART_AP_SIZE_SHFT);	/* 16MB */
+		break;
+	case 32:
+		ap_reg |= (0x3f8 << CA_GART_AP_SIZE_SHFT);	/* 32 MB */
+		break;
+	case 64:
+		ap_reg |= (0x3f0 << CA_GART_AP_SIZE_SHFT);	/* 64 MB */
+		break;
+	case 128:
+		ap_reg |= (0x3e0 << CA_GART_AP_SIZE_SHFT);	/* 128 MB */
+		break;
+	case 256:
+		ap_reg |= (0x3c0 << CA_GART_AP_SIZE_SHFT);	/* 256 MB */
+		break;
+	case 512:
+		ap_reg |= (0x380 << CA_GART_AP_SIZE_SHFT);	/* 512 MB */
+		break;
+	case 1024:
+		ap_reg |= (0x300 << CA_GART_AP_SIZE_SHFT);	/* 1GB */
+		break;
+	case 2048:
+		ap_reg |= (0x200 << CA_GART_AP_SIZE_SHFT);	/* 2GB */
+		break;
+	case 4096:
+		ap_reg |= (0x000 << CA_GART_AP_SIZE_SHFT);	/* 4 GB */
+		break;
+	default:
+		printk(KERN_ERR "%s:  Invalid CA_APERATURE_SIZE "
+		       "0x%lx\n", __func__, (ulong) CA_APERATURE_SIZE);
+		return -1;
+	}
+
+	/*
+	 * Set up other aperature parameters
+	 */
+
+	if (PAGE_SIZE >= 16384) {
+		tioca_kern->ca_ap_pagesize = 16384;
+		ap_reg |= CA_GART_PAGE_SIZE;
+	} else {
+		tioca_kern->ca_ap_pagesize = 4096;
+	}
+
+	tioca_kern->ca_ap_size = CA_APERATURE_SIZE;
+	tioca_kern->ca_ap_bus_base = CA_APERATURE_BASE;
+	tioca_kern->ca_gart_entries =
+	    tioca_kern->ca_ap_size / tioca_kern->ca_ap_pagesize;
+
+	ap_reg |= (CA_GART_AP_ENB_AGP | CA_GART_AP_ENB_PCI);
+	ap_reg |= tioca_kern->ca_ap_bus_base;
+
+	/*
+	 * Allocate and set up the GART
+	 */
+
+	tioca_kern->ca_gart_size = tioca_kern->ca_gart_entries * sizeof(u64);
+	tmp =
+	    alloc_pages_node(tioca_kern->ca_closest_node,
+			     GFP_KERNEL | __GFP_ZERO,
+			     get_order(tioca_kern->ca_gart_size));
+
+	if (!tmp) {
+		printk(KERN_ERR "%s:  Could not allocate "
+		       "%llu bytes (order %d) for GART\n",
+		       __func__,
+		       tioca_kern->ca_gart_size,
+		       get_order(tioca_kern->ca_gart_size));
+		return -ENOMEM;
+	}
+
+	tioca_kern->ca_gart = page_address(tmp);
+	tioca_kern->ca_gart_coretalk_addr =
+	    PHYS_TO_TIODMA(virt_to_phys(tioca_kern->ca_gart));
+
+	/*
+	 * Compute PCI/AGP convenience fields 
+	 */
+
+	offset = CA_PCI32_MAPPED_BASE - CA_APERATURE_BASE;
+	tioca_kern->ca_pciap_base = CA_PCI32_MAPPED_BASE;
+	tioca_kern->ca_pciap_size = CA_PCI32_MAPPED_SIZE;
+	tioca_kern->ca_pcigart_start = offset / tioca_kern->ca_ap_pagesize;
+	tioca_kern->ca_pcigart_base =
+	    tioca_kern->ca_gart_coretalk_addr + offset;
+	tioca_kern->ca_pcigart =
+	    &tioca_kern->ca_gart[tioca_kern->ca_pcigart_start];
+	tioca_kern->ca_pcigart_entries =
+	    tioca_kern->ca_pciap_size / tioca_kern->ca_ap_pagesize;
+	tioca_kern->ca_pcigart_pagemap =
+	    kzalloc(tioca_kern->ca_pcigart_entries / 8, GFP_KERNEL);
+	if (!tioca_kern->ca_pcigart_pagemap) {
+		free_pages((unsigned long)tioca_kern->ca_gart,
+			   get_order(tioca_kern->ca_gart_size));
+		return -1;
+	}
+
+	offset = CA_AGP_MAPPED_BASE - CA_APERATURE_BASE;
+	tioca_kern->ca_gfxap_base = CA_AGP_MAPPED_BASE;
+	tioca_kern->ca_gfxap_size = CA_AGP_MAPPED_SIZE;
+	tioca_kern->ca_gfxgart_start = offset / tioca_kern->ca_ap_pagesize;
+	tioca_kern->ca_gfxgart_base =
+	    tioca_kern->ca_gart_coretalk_addr + offset;
+	tioca_kern->ca_gfxgart =
+	    &tioca_kern->ca_gart[tioca_kern->ca_gfxgart_start];
+	tioca_kern->ca_gfxgart_entries =
+	    tioca_kern->ca_gfxap_size / tioca_kern->ca_ap_pagesize;
+
+	/*
+	 * various control settings:
+	 *      use agp op-combining
+	 *      use GET semantics to fetch memory
+	 *      participate in coherency domain
+	 * 	DISABLE GART PREFETCHING due to hw bug tracked in SGI PV930029
+	 */
+
+	__sn_setq_relaxed(&ca_base->ca_control1,
+			CA_AGPDMA_OP_ENB_COMBDELAY);	/* PV895469 ? */
+	__sn_clrq_relaxed(&ca_base->ca_control2, CA_GART_MEM_PARAM);
+	__sn_setq_relaxed(&ca_base->ca_control2,
+			(0x2ull << CA_GART_MEM_PARAM_SHFT));
+	tioca_kern->ca_gart_iscoherent = 1;
+	__sn_clrq_relaxed(&ca_base->ca_control2,
+	    		(CA_GART_WR_PREFETCH_ENB | CA_GART_RD_PREFETCH_ENB));
+
+	/*
+	 * Unmask GART fetch error interrupts.  Clear residual errors first.
+	 */
+
+	writeq(CA_GART_FETCH_ERR, &ca_base->ca_int_status_alias);
+	writeq(CA_GART_FETCH_ERR, &ca_base->ca_mult_error_alias);
+	__sn_clrq_relaxed(&ca_base->ca_int_mask, CA_GART_FETCH_ERR);
+
+	/*
+	 * Program the aperature and gart registers in TIOCA
+	 */
+
+	writeq(ap_reg, &ca_base->ca_gart_aperature);
+	writeq(tioca_kern->ca_gart_coretalk_addr|1, &ca_base->ca_gart_ptr_table);
+
+	return 0;
+}
+
+/**
+ * tioca_fastwrite_enable - enable AGP FW for a tioca and its functions
+ * @tioca_kernel: structure representing the CA
+ *
+ * Given a CA, scan all attached functions making sure they all support
+ * FastWrite.  If so, enable FastWrite for all functions and the CA itself.
+ */
+
+void
+tioca_fastwrite_enable(struct tioca_kernel *tioca_kern)
+{
+	int cap_ptr;
+	u32 reg;
+	struct tioca __iomem *tioca_base;
+	struct pci_dev *pdev;
+	struct tioca_common *common;
+
+	common = tioca_kern->ca_common;
+
+	/*
+	 * Scan all vga controllers on this bus making sure they all
+	 * support FW.  If not, return.
+	 */
+
+	list_for_each_entry(pdev, tioca_kern->ca_devices, bus_list) {
+		if (pdev->class != (PCI_CLASS_DISPLAY_VGA << 8))
+			continue;
+
+		cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
+		if (!cap_ptr)
+			return;	/* no AGP CAP means no FW */
+
+		pci_read_config_dword(pdev, cap_ptr + PCI_AGP_STATUS, &reg);
+		if (!(reg & PCI_AGP_STATUS_FW))
+			return;	/* function doesn't support FW */
+	}
+
+	/*
+	 * Set fw for all vga fn's
+	 */
+
+	list_for_each_entry(pdev, tioca_kern->ca_devices, bus_list) {
+		if (pdev->class != (PCI_CLASS_DISPLAY_VGA << 8))
+			continue;
+
+		cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
+		pci_read_config_dword(pdev, cap_ptr + PCI_AGP_COMMAND, &reg);
+		reg |= PCI_AGP_COMMAND_FW;
+		pci_write_config_dword(pdev, cap_ptr + PCI_AGP_COMMAND, reg);
+	}
+
+	/*
+	 * Set ca's fw to match
+	 */
+
+	tioca_base = (struct tioca __iomem*)common->ca_common.bs_base;
+	__sn_setq_relaxed(&tioca_base->ca_control1, CA_AGP_FW_ENABLE);
+}
+
+EXPORT_SYMBOL(tioca_fastwrite_enable);	/* used by agp-sgi */
+
+/**
+ * tioca_dma_d64 - create a DMA mapping using 64-bit direct mode
+ * @paddr: system physical address
+ *
+ * Map @paddr into 64-bit CA bus space.  No device context is necessary.
+ * Bits 53:0 come from the coretalk address.  We just need to mask in the
+ * following optional bits of the 64-bit pci address:
+ *
+ * 63:60 - Coretalk Packet Type -  0x1 for Mem Get/Put (coherent)
+ *                                 0x2 for PIO (non-coherent)
+ *                                 We will always use 0x1
+ * 55:55 - Swap bytes		   Currently unused
+ */
+static u64
+tioca_dma_d64(unsigned long paddr)
+{
+	dma_addr_t bus_addr;
+
+	bus_addr = PHYS_TO_TIODMA(paddr);
+
+	BUG_ON(!bus_addr);
+	BUG_ON(bus_addr >> 54);
+
+	/* Set upper nibble to Cache Coherent Memory op */
+	bus_addr |= (1UL << 60);
+
+	return bus_addr;
+}
+
+/**
+ * tioca_dma_d48 - create a DMA mapping using 48-bit direct mode
+ * @pdev: linux pci_dev representing the function
+ * @paddr: system physical address
+ *
+ * Map @paddr into 64-bit bus space of the CA associated with @pcidev_info.
+ *
+ * The CA agp 48 bit direct address falls out as follows:
+ *
+ * When direct mapping AGP addresses, the 48 bit AGP address is
+ * constructed as follows:
+ *
+ * [47:40] - Low 8 bits of the page Node ID extracted from coretalk
+ *              address [47:40].  The upper 8 node bits are fixed
+ *              and come from the xxx register bits [5:0]
+ * [39:38] - Chiplet ID extracted from coretalk address [39:38]
+ * [37:00] - node offset extracted from coretalk address [37:00]
+ * 
+ * Since the node id in general will be non-zero, and the chiplet id
+ * will always be non-zero, it follows that the device must support
+ * a dma mask of at least 0xffffffffff (40 bits) to target node 0
+ * and in general should be 0xffffffffffff (48 bits) to target nodes
+ * up to 255.  Nodes above 255 need the support of the xxx register,
+ * and so a given CA can only directly target nodes in the range
+ * xxx - xxx+255.
+ */
+static u64
+tioca_dma_d48(struct pci_dev *pdev, u64 paddr)
+{
+	struct tioca_common *tioca_common;
+	struct tioca __iomem *ca_base;
+	u64 ct_addr;
+	dma_addr_t bus_addr;
+	u32 node_upper;
+	u64 agp_dma_extn;
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev);
+
+	tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info;
+	ca_base = (struct tioca __iomem *)tioca_common->ca_common.bs_base;
+
+	ct_addr = PHYS_TO_TIODMA(paddr);
+	if (!ct_addr)
+		return 0;
+
+	bus_addr = (dma_addr_t) (ct_addr & 0xffffffffffffUL);
+	node_upper = ct_addr >> 48;
+
+	if (node_upper > 64) {
+		printk(KERN_ERR "%s:  coretalk addr 0x%p node id out "
+		       "of range\n", __func__, (void *)ct_addr);
+		return 0;
+	}
+
+	agp_dma_extn = __sn_readq_relaxed(&ca_base->ca_agp_dma_addr_extn);
+	if (node_upper != (agp_dma_extn >> CA_AGP_DMA_NODE_ID_SHFT)) {
+		printk(KERN_ERR "%s:  coretalk upper node (%u) "
+		       "mismatch with ca_agp_dma_addr_extn (%llu)\n",
+		       __func__,
+		       node_upper, (agp_dma_extn >> CA_AGP_DMA_NODE_ID_SHFT));
+		return 0;
+	}
+
+	return bus_addr;
+}
+
+/**
+ * tioca_dma_mapped - create a DMA mapping using a CA GART 
+ * @pdev: linux pci_dev representing the function
+ * @paddr: host physical address to map
+ * @req_size: len (bytes) to map
+ *
+ * Map @paddr into CA address space using the GART mechanism.  The mapped
+ * dma_addr_t is guaranteed to be contiguous in CA bus space.
+ */
+static dma_addr_t
+tioca_dma_mapped(struct pci_dev *pdev, unsigned long paddr, size_t req_size)
+{
+	int ps, ps_shift, entry, entries, mapsize;
+	u64 xio_addr, end_xio_addr;
+	struct tioca_common *tioca_common;
+	struct tioca_kernel *tioca_kern;
+	dma_addr_t bus_addr = 0;
+	struct tioca_dmamap *ca_dmamap;
+	void *map;
+	unsigned long flags;
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev);
+
+	tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info;
+	tioca_kern = (struct tioca_kernel *)tioca_common->ca_kernel_private;
+
+	xio_addr = PHYS_TO_TIODMA(paddr);
+	if (!xio_addr)
+		return 0;
+
+	spin_lock_irqsave(&tioca_kern->ca_lock, flags);
+
+	/*
+	 * allocate a map struct
+	 */
+
+	ca_dmamap = kzalloc(sizeof(struct tioca_dmamap), GFP_ATOMIC);
+	if (!ca_dmamap)
+		goto map_return;
+
+	/*
+	 * Locate free entries that can hold req_size.  Account for
+	 * unaligned start/length when allocating.
+	 */
+
+	ps = tioca_kern->ca_ap_pagesize;	/* will be power of 2 */
+	ps_shift = ffs(ps) - 1;
+	end_xio_addr = xio_addr + req_size - 1;
+
+	entries = (end_xio_addr >> ps_shift) - (xio_addr >> ps_shift) + 1;
+
+	map = tioca_kern->ca_pcigart_pagemap;
+	mapsize = tioca_kern->ca_pcigart_entries;
+
+	entry = bitmap_find_next_zero_area(map, mapsize, 0, entries, 0);
+	if (entry >= mapsize) {
+		kfree(ca_dmamap);
+		goto map_return;
+	}
+
+	bitmap_set(map, entry, entries);
+
+	bus_addr = tioca_kern->ca_pciap_base + (entry * ps);
+
+	ca_dmamap->cad_dma_addr = bus_addr;
+	ca_dmamap->cad_gart_size = entries;
+	ca_dmamap->cad_gart_entry = entry;
+	list_add(&ca_dmamap->cad_list, &tioca_kern->ca_dmamaps);
+
+	if (xio_addr % ps) {
+		tioca_kern->ca_pcigart[entry] = tioca_paddr_to_gart(xio_addr);
+		bus_addr += xio_addr & (ps - 1);
+		xio_addr &= ~(ps - 1);
+		xio_addr += ps;
+		entry++;
+	}
+
+	while (xio_addr < end_xio_addr) {
+		tioca_kern->ca_pcigart[entry] = tioca_paddr_to_gart(xio_addr);
+		xio_addr += ps;
+		entry++;
+	}
+
+	tioca_tlbflush(tioca_kern);
+
+map_return:
+	spin_unlock_irqrestore(&tioca_kern->ca_lock, flags);
+	return bus_addr;
+}
+
+/**
+ * tioca_dma_unmap - release CA mapping resources
+ * @pdev: linux pci_dev representing the function
+ * @bus_addr: bus address returned by an earlier tioca_dma_map
+ * @dir: mapping direction (unused)
+ *
+ * Locate mapping resources associated with @bus_addr and release them.
+ * For mappings created using the direct modes (64 or 48) there are no
+ * resources to release.
+ */
+static void
+tioca_dma_unmap(struct pci_dev *pdev, dma_addr_t bus_addr, int dir)
+{
+	int i, entry;
+	struct tioca_common *tioca_common;
+	struct tioca_kernel *tioca_kern;
+	struct tioca_dmamap *map;
+	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev);
+	unsigned long flags;
+
+	tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info;
+	tioca_kern = (struct tioca_kernel *)tioca_common->ca_kernel_private;
+
+	/* return straight away if this isn't be a mapped address */
+
+	if (bus_addr < tioca_kern->ca_pciap_base ||
+	    bus_addr >= (tioca_kern->ca_pciap_base + tioca_kern->ca_pciap_size))
+		return;
+
+	spin_lock_irqsave(&tioca_kern->ca_lock, flags);
+
+	list_for_each_entry(map, &tioca_kern->ca_dmamaps, cad_list)
+	    if (map->cad_dma_addr == bus_addr)
+		break;
+
+	BUG_ON(map == NULL);
+
+	entry = map->cad_gart_entry;
+
+	for (i = 0; i < map->cad_gart_size; i++, entry++) {
+		clear_bit(entry, tioca_kern->ca_pcigart_pagemap);
+		tioca_kern->ca_pcigart[entry] = 0;
+	}
+	tioca_tlbflush(tioca_kern);
+
+	list_del(&map->cad_list);
+	spin_unlock_irqrestore(&tioca_kern->ca_lock, flags);
+	kfree(map);
+}
+
+/**
+ * tioca_dma_map - map pages for PCI DMA
+ * @pdev: linux pci_dev representing the function
+ * @paddr: host physical address to map
+ * @byte_count: bytes to map
+ *
+ * This is the main wrapper for mapping host physical pages to CA PCI space.
+ * The mapping mode used is based on the devices dma_mask.  As a last resort
+ * use the GART mapped mode.
+ */
+static u64
+tioca_dma_map(struct pci_dev *pdev, unsigned long paddr, size_t byte_count, int dma_flags)
+{
+	u64 mapaddr;
+
+	/*
+	 * Not supported for now ...
+	 */
+	if (dma_flags & SN_DMA_MSI)
+		return 0;
+
+	/*
+	 * If card is 64 or 48 bit addressable, use a direct mapping.  32
+	 * bit direct is so restrictive w.r.t. where the memory resides that
+	 * we don't use it even though CA has some support.
+	 */
+
+	if (pdev->dma_mask == ~0UL)
+		mapaddr = tioca_dma_d64(paddr);
+	else if (pdev->dma_mask == 0xffffffffffffUL)
+		mapaddr = tioca_dma_d48(pdev, paddr);
+	else
+		mapaddr = 0;
+
+	/* Last resort ... use PCI portion of CA GART */
+
+	if (mapaddr == 0)
+		mapaddr = tioca_dma_mapped(pdev, paddr, byte_count);
+
+	return mapaddr;
+}
+
+/**
+ * tioca_error_intr_handler - SGI TIO CA error interrupt handler
+ * @irq: unused
+ * @arg: pointer to tioca_common struct for the given CA
+ *
+ * Handle a CA error interrupt.  Simply a wrapper around a SAL call which
+ * defers processing to the SGI prom.
+ */
+static irqreturn_t
+tioca_error_intr_handler(int irq, void *arg)
+{
+	struct tioca_common *soft = arg;
+	struct ia64_sal_retval ret_stuff;
+	u64 segment;
+	u64 busnum;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	segment = soft->ca_common.bs_persist_segment;
+	busnum = soft->ca_common.bs_persist_busnum;
+
+	SAL_CALL_NOLOCK(ret_stuff,
+			(u64) SN_SAL_IOIF_ERROR_INTERRUPT,
+			segment, busnum, 0, 0, 0, 0, 0);
+
+	return IRQ_HANDLED;
+}
+
+/**
+ * tioca_bus_fixup - perform final PCI fixup for a TIO CA bus
+ * @prom_bussoft: Common prom/kernel struct representing the bus
+ *
+ * Replicates the tioca_common pointed to by @prom_bussoft in kernel
+ * space.  Allocates and initializes a kernel-only area for a given CA,
+ * and sets up an irq for handling CA error interrupts.
+ *
+ * On successful setup, returns the kernel version of tioca_common back to
+ * the caller.
+ */
+static void *
+tioca_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller)
+{
+	struct tioca_common *tioca_common;
+	struct tioca_kernel *tioca_kern;
+	struct pci_bus *bus;
+
+	/* sanity check prom rev */
+
+	if (is_shub1() && sn_sal_rev() < 0x0406) {
+		printk
+		    (KERN_ERR "%s:  SGI prom rev 4.06 or greater required "
+		     "for tioca support\n", __func__);
+		return NULL;
+	}
+
+	/*
+	 * Allocate kernel bus soft and copy from prom.
+	 */
+
+	tioca_common = kmemdup(prom_bussoft, sizeof(struct tioca_common),
+			       GFP_KERNEL);
+	if (!tioca_common)
+		return NULL;
+
+	tioca_common->ca_common.bs_base = (unsigned long)
+		ioremap(REGION_OFFSET(tioca_common->ca_common.bs_base),
+			sizeof(struct tioca_common));
+
+	/* init kernel-private area */
+
+	tioca_kern = kzalloc(sizeof(struct tioca_kernel), GFP_KERNEL);
+	if (!tioca_kern) {
+		kfree(tioca_common);
+		return NULL;
+	}
+
+	tioca_kern->ca_common = tioca_common;
+	spin_lock_init(&tioca_kern->ca_lock);
+	INIT_LIST_HEAD(&tioca_kern->ca_dmamaps);
+	tioca_kern->ca_closest_node =
+	    nasid_to_cnodeid(tioca_common->ca_closest_nasid);
+	tioca_common->ca_kernel_private = (u64) tioca_kern;
+
+	bus = pci_find_bus(tioca_common->ca_common.bs_persist_segment,
+		tioca_common->ca_common.bs_persist_busnum);
+	BUG_ON(!bus);
+	tioca_kern->ca_devices = &bus->devices;
+
+	/* init GART */
+
+	if (tioca_gart_init(tioca_kern) < 0) {
+		kfree(tioca_kern);
+		kfree(tioca_common);
+		return NULL;
+	}
+
+	tioca_gart_found++;
+	list_add(&tioca_kern->ca_list, &tioca_list);
+
+	if (request_irq(SGI_TIOCA_ERROR,
+			tioca_error_intr_handler,
+			IRQF_SHARED, "TIOCA error", (void *)tioca_common))
+		printk(KERN_WARNING
+		       "%s:  Unable to get irq %d.  "
+		       "Error interrupts won't be routed for TIOCA bus %d\n",
+		       __func__, SGI_TIOCA_ERROR,
+		       (int)tioca_common->ca_common.bs_persist_busnum);
+
+	irq_set_handler(SGI_TIOCA_ERROR, handle_level_irq);
+	sn_set_err_irq_affinity(SGI_TIOCA_ERROR);
+
+	/* Setup locality information */
+	controller->node = tioca_kern->ca_closest_node;
+	return tioca_common;
+}
+
+static struct sn_pcibus_provider tioca_pci_interfaces = {
+	.dma_map = tioca_dma_map,
+	.dma_map_consistent = tioca_dma_map,
+	.dma_unmap = tioca_dma_unmap,
+	.bus_fixup = tioca_bus_fixup,
+	.force_interrupt = NULL,
+	.target_interrupt = NULL
+};
+
+/**
+ * tioca_init_provider - init SN PCI provider ops for TIO CA
+ */
+int
+tioca_init_provider(void)
+{
+	sn_pci_provider[PCIIO_ASIC_TYPE_TIOCA] = &tioca_pci_interfaces;
+	return 0;
+}
diff --git a/kernel/arch/ia64/sn/pci/tioce_provider.c b/kernel/arch/ia64/sn/pci/tioce_provider.c
new file mode 100644
index 000000000..46d3df4b0
--- /dev/null
+++ b/kernel/arch/ia64/sn/pci/tioce_provider.c
@@ -0,0 +1,1062 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2003-2006 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/types.h>
+#include <linux/interrupt.h>
+#include <linux/slab.h>
+#include <linux/pci.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/sn/addrs.h>
+#include <asm/sn/io.h>
+#include <asm/sn/pcidev.h>
+#include <asm/sn/pcibus_provider_defs.h>
+#include <asm/sn/tioce_provider.h>
+
+/*
+ * 1/26/2006
+ *
+ * WAR for SGI PV 944642.  For revA TIOCE, need to use the following recipe
+ * (taken from the above PV) before and after accessing tioce internal MMR's
+ * to avoid tioce lockups.
+ *
+ * The recipe as taken from the PV:
+ *
+ *	if(mmr address < 0x45000) {
+ *		if(mmr address == 0 or 0x80)
+ *			mmr wrt or read address 0xc0
+ *		else if(mmr address == 0x148 or 0x200)
+ *			mmr wrt or read address 0x28
+ *		else
+ *			mmr wrt or read address 0x158
+ *
+ *		do desired mmr access (rd or wrt)
+ *
+ *		if(mmr address == 0x100)
+ *			mmr wrt or read address 0x38
+ *		mmr wrt or read address 0xb050
+ *	} else
+ *		do desired mmr access
+ *
+ * According to hw, we can use reads instead of writes to the above address
+ *
+ * Note this WAR can only to be used for accessing internal MMR's in the
+ * TIOCE Coretalk Address Range 0x0 - 0x07ff_ffff.  This includes the
+ * "Local CE Registers and Memories" and "PCI Compatible Config Space" address
+ * spaces from table 2-1 of the "CE Programmer's Reference Overview" document.
+ *
+ * All registers defined in struct tioce will meet that criteria.
+ */
+
+static void inline
+tioce_mmr_war_pre(struct tioce_kernel *kern, void __iomem *mmr_addr)
+{
+	u64 mmr_base;
+	u64 mmr_offset;
+
+	if (kern->ce_common->ce_rev != TIOCE_REV_A)
+		return;
+
+	mmr_base = kern->ce_common->ce_pcibus.bs_base;
+	mmr_offset = (unsigned long)mmr_addr - mmr_base;
+
+	if (mmr_offset < 0x45000) {
+		u64 mmr_war_offset;
+
+		if (mmr_offset == 0 || mmr_offset == 0x80)
+			mmr_war_offset = 0xc0;
+		else if (mmr_offset == 0x148 || mmr_offset == 0x200)
+			mmr_war_offset = 0x28;
+		else
+			mmr_war_offset = 0x158;
+
+		readq_relaxed((void __iomem *)(mmr_base + mmr_war_offset));
+	}
+}
+
+static void inline
+tioce_mmr_war_post(struct tioce_kernel *kern, void __iomem *mmr_addr)
+{
+	u64 mmr_base;
+	u64 mmr_offset;
+
+	if (kern->ce_common->ce_rev != TIOCE_REV_A)
+		return;
+
+	mmr_base = kern->ce_common->ce_pcibus.bs_base;
+	mmr_offset = (unsigned long)mmr_addr - mmr_base;
+
+	if (mmr_offset < 0x45000) {
+		if (mmr_offset == 0x100)
+			readq_relaxed((void __iomem *)(mmr_base + 0x38));
+		readq_relaxed((void __iomem *)(mmr_base + 0xb050));
+	}
+}
+
+/* load mmr contents into a variable */
+#define tioce_mmr_load(kern, mmrp, varp) do {\
+	tioce_mmr_war_pre(kern, mmrp); \
+	*(varp) = readq_relaxed(mmrp); \
+	tioce_mmr_war_post(kern, mmrp); \
+} while (0)
+
+/* store variable contents into mmr */
+#define tioce_mmr_store(kern, mmrp, varp) do {\
+	tioce_mmr_war_pre(kern, mmrp); \
+	writeq(*varp, mmrp); \
+	tioce_mmr_war_post(kern, mmrp); \
+} while (0)
+
+/* store immediate value into mmr */
+#define tioce_mmr_storei(kern, mmrp, val) do {\
+	tioce_mmr_war_pre(kern, mmrp); \
+	writeq(val, mmrp); \
+	tioce_mmr_war_post(kern, mmrp); \
+} while (0)
+
+/* set bits (immediate value) into mmr */
+#define tioce_mmr_seti(kern, mmrp, bits) do {\
+	u64 tmp; \
+	tioce_mmr_load(kern, mmrp, &tmp); \
+	tmp |= (bits); \
+	tioce_mmr_store(kern, mmrp, &tmp); \
+} while (0)
+
+/* clear bits (immediate value) into mmr */
+#define tioce_mmr_clri(kern, mmrp, bits) do { \
+	u64 tmp; \
+	tioce_mmr_load(kern, mmrp, &tmp); \
+	tmp &= ~(bits); \
+	tioce_mmr_store(kern, mmrp, &tmp); \
+} while (0)
+
+/**
+ * Bus address ranges for the 5 flavors of TIOCE DMA
+ */
+
+#define TIOCE_D64_MIN	0x8000000000000000UL
+#define TIOCE_D64_MAX	0xffffffffffffffffUL
+#define TIOCE_D64_ADDR(a)	((a) >= TIOCE_D64_MIN)
+
+#define TIOCE_D32_MIN	0x0000000080000000UL
+#define TIOCE_D32_MAX	0x00000000ffffffffUL
+#define TIOCE_D32_ADDR(a)	((a) >= TIOCE_D32_MIN && (a) <= TIOCE_D32_MAX)
+
+#define TIOCE_M32_MIN	0x0000000000000000UL
+#define TIOCE_M32_MAX	0x000000007fffffffUL
+#define TIOCE_M32_ADDR(a)	((a) >= TIOCE_M32_MIN && (a) <= TIOCE_M32_MAX)
+
+#define TIOCE_M40_MIN	0x0000004000000000UL
+#define TIOCE_M40_MAX	0x0000007fffffffffUL
+#define TIOCE_M40_ADDR(a)	((a) >= TIOCE_M40_MIN && (a) <= TIOCE_M40_MAX)
+
+#define TIOCE_M40S_MIN	0x0000008000000000UL
+#define TIOCE_M40S_MAX	0x000000ffffffffffUL
+#define TIOCE_M40S_ADDR(a)	((a) >= TIOCE_M40S_MIN && (a) <= TIOCE_M40S_MAX)
+
+/*
+ * ATE manipulation macros.
+ */
+
+#define ATE_PAGESHIFT(ps)	(__ffs(ps))
+#define ATE_PAGEMASK(ps)	((ps)-1)
+
+#define ATE_PAGE(x, ps) ((x) >> ATE_PAGESHIFT(ps))
+#define ATE_NPAGES(start, len, pagesize) \
+	(ATE_PAGE((start)+(len)-1, pagesize) - ATE_PAGE(start, pagesize) + 1)
+
+#define ATE_VALID(ate)	((ate) & (1UL << 63))
+#define ATE_MAKE(addr, ps, msi) \
+	(((addr) & ~ATE_PAGEMASK(ps)) | (1UL << 63) | ((msi)?(1UL << 62):0))
+
+/*
+ * Flavors of ate-based mapping supported by tioce_alloc_map()
+ */
+
+#define TIOCE_ATE_M32	1
+#define TIOCE_ATE_M40	2
+#define TIOCE_ATE_M40S	3
+
+#define KB(x)	((u64)(x) << 10)
+#define MB(x)	((u64)(x) << 20)
+#define GB(x)	((u64)(x) << 30)
+
+/**
+ * tioce_dma_d64 - create a DMA mapping using 64-bit direct mode
+ * @ct_addr: system coretalk address
+ *
+ * Map @ct_addr into 64-bit CE bus space.  No device context is necessary
+ * and no CE mapping are consumed.
+ *
+ * Bits 53:0 come from the coretalk address.  The remaining bits are set as
+ * follows:
+ *
+ * 63    - must be 1 to indicate d64 mode to CE hardware
+ * 62    - barrier bit ... controlled with tioce_dma_barrier()
+ * 61    - msi bit ... specified through dma_flags
+ * 60:54 - reserved, MBZ
+ */
+static u64
+tioce_dma_d64(unsigned long ct_addr, int dma_flags)
+{
+	u64 bus_addr;
+
+	bus_addr = ct_addr | (1UL << 63);
+	if (dma_flags & SN_DMA_MSI)
+		bus_addr |= (1UL << 61);
+
+	return bus_addr;
+}
+
+/**
+ * pcidev_to_tioce - return misc ce related pointers given a pci_dev
+ * @pci_dev: pci device context
+ * @base: ptr to store struct tioce_mmr * for the CE holding this device
+ * @kernel: ptr to store struct tioce_kernel * for the CE holding this device
+ * @port: ptr to store the CE port number that this device is on
+ *
+ * Return pointers to various CE-related structures for the CE upstream of
+ * @pci_dev.
+ */
+static inline void
+pcidev_to_tioce(struct pci_dev *pdev, struct tioce __iomem **base,
+		struct tioce_kernel **kernel, int *port)
+{
+	struct pcidev_info *pcidev_info;
+	struct tioce_common *ce_common;
+	struct tioce_kernel *ce_kernel;
+
+	pcidev_info = SN_PCIDEV_INFO(pdev);
+	ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info;
+	ce_kernel = (struct tioce_kernel *)ce_common->ce_kernel_private;
+
+	if (base)
+		*base = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base;
+	if (kernel)
+		*kernel = ce_kernel;
+
+	/*
+	 * we use port as a zero-based value internally, even though the
+	 * documentation is 1-based.
+	 */
+	if (port)
+		*port =
+		    (pdev->bus->number < ce_kernel->ce_port1_secondary) ? 0 : 1;
+}
+
+/**
+ * tioce_alloc_map - Given a coretalk address, map it to pcie bus address
+ * space using one of the various ATE-based address modes.
+ * @ce_kern: tioce context
+ * @type: map mode to use
+ * @port: 0-based port that the requesting device is downstream of
+ * @ct_addr: the coretalk address to map
+ * @len: number of bytes to map
+ *
+ * Given the addressing type, set up various parameters that define the
+ * ATE pool to use.  Search for a contiguous block of entries to cover the
+ * length, and if enough resources exist, fill in the ATEs and construct a
+ * tioce_dmamap struct to track the mapping.
+ */
+static u64
+tioce_alloc_map(struct tioce_kernel *ce_kern, int type, int port,
+		u64 ct_addr, int len, int dma_flags)
+{
+	int i;
+	int j;
+	int first;
+	int last;
+	int entries;
+	int nates;
+	u64 pagesize;
+	int msi_capable, msi_wanted;
+	u64 *ate_shadow;
+	u64 __iomem *ate_reg;
+	u64 addr;
+	struct tioce __iomem *ce_mmr;
+	u64 bus_base;
+	struct tioce_dmamap *map;
+
+	ce_mmr = (struct tioce __iomem *)ce_kern->ce_common->ce_pcibus.bs_base;
+
+	switch (type) {
+	case TIOCE_ATE_M32:
+		/*
+		 * The first 64 entries of the ate3240 pool are dedicated to
+		 * super-page (TIOCE_ATE_M40S) mode.
+		 */
+		first = 64;
+		entries = TIOCE_NUM_M3240_ATES - 64;
+		ate_shadow = ce_kern->ce_ate3240_shadow;
+		ate_reg = ce_mmr->ce_ure_ate3240;
+		pagesize = ce_kern->ce_ate3240_pagesize;
+		bus_base = TIOCE_M32_MIN;
+		msi_capable = 1;
+		break;
+	case TIOCE_ATE_M40:
+		first = 0;
+		entries = TIOCE_NUM_M40_ATES;
+		ate_shadow = ce_kern->ce_ate40_shadow;
+		ate_reg = ce_mmr->ce_ure_ate40;
+		pagesize = MB(64);
+		bus_base = TIOCE_M40_MIN;
+		msi_capable = 0;
+		break;
+	case TIOCE_ATE_M40S:
+		/*
+		 * ate3240 entries 0-31 are dedicated to port1 super-page
+		 * mappings.  ate3240 entries 32-63 are dedicated to port2.
+		 */
+		first = port * 32;
+		entries = 32;
+		ate_shadow = ce_kern->ce_ate3240_shadow;
+		ate_reg = ce_mmr->ce_ure_ate3240;
+		pagesize = GB(16);
+		bus_base = TIOCE_M40S_MIN;
+		msi_capable = 0;
+		break;
+	default:
+		return 0;
+	}
+
+	msi_wanted = dma_flags & SN_DMA_MSI;
+	if (msi_wanted && !msi_capable)
+		return 0;
+
+	nates = ATE_NPAGES(ct_addr, len, pagesize);
+	if (nates > entries)
+		return 0;
+
+	last = first + entries - nates;
+	for (i = first; i <= last; i++) {
+		if (ATE_VALID(ate_shadow[i]))
+			continue;
+
+		for (j = i; j < i + nates; j++)
+			if (ATE_VALID(ate_shadow[j]))
+				break;
+
+		if (j >= i + nates)
+			break;
+	}
+
+	if (i > last)
+		return 0;
+
+	map = kzalloc(sizeof(struct tioce_dmamap), GFP_ATOMIC);
+	if (!map)
+		return 0;
+
+	addr = ct_addr;
+	for (j = 0; j < nates; j++) {
+		u64 ate;
+
+		ate = ATE_MAKE(addr, pagesize, msi_wanted);
+		ate_shadow[i + j] = ate;
+		tioce_mmr_storei(ce_kern, &ate_reg[i + j], ate);
+		addr += pagesize;
+	}
+
+	map->refcnt = 1;
+	map->nbytes = nates * pagesize;
+	map->ct_start = ct_addr & ~ATE_PAGEMASK(pagesize);
+	map->pci_start = bus_base + (i * pagesize);
+	map->ate_hw = &ate_reg[i];
+	map->ate_shadow = &ate_shadow[i];
+	map->ate_count = nates;
+
+	list_add(&map->ce_dmamap_list, &ce_kern->ce_dmamap_list);
+
+	return (map->pci_start + (ct_addr - map->ct_start));
+}
+
+/**
+ * tioce_dma_d32 - create a DMA mapping using 32-bit direct mode
+ * @pdev: linux pci_dev representing the function
+ * @paddr: system physical address
+ *
+ * Map @paddr into 32-bit bus space of the CE associated with @pcidev_info.
+ */
+static u64
+tioce_dma_d32(struct pci_dev *pdev, u64 ct_addr, int dma_flags)
+{
+	int dma_ok;
+	int port;
+	struct tioce __iomem *ce_mmr;
+	struct tioce_kernel *ce_kern;
+	u64 ct_upper;
+	u64 ct_lower;
+	dma_addr_t bus_addr;
+
+	if (dma_flags & SN_DMA_MSI)
+		return 0;
+
+	ct_upper = ct_addr & ~0x3fffffffUL;
+	ct_lower = ct_addr & 0x3fffffffUL;
+
+	pcidev_to_tioce(pdev, &ce_mmr, &ce_kern, &port);
+
+	if (ce_kern->ce_port[port].dirmap_refcnt == 0) {
+		u64 tmp;
+
+		ce_kern->ce_port[port].dirmap_shadow = ct_upper;
+		tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_dir_map[port],
+				 ct_upper);
+		tmp = ce_mmr->ce_ure_dir_map[port];
+		dma_ok = 1;
+	} else
+		dma_ok = (ce_kern->ce_port[port].dirmap_shadow == ct_upper);
+
+	if (dma_ok) {
+		ce_kern->ce_port[port].dirmap_refcnt++;
+		bus_addr = TIOCE_D32_MIN + ct_lower;
+	} else
+		bus_addr = 0;
+
+	return bus_addr;
+}
+
+/**
+ * tioce_dma_barrier - swizzle a TIOCE bus address to include or exclude
+ * the barrier bit.
+ * @bus_addr:  bus address to swizzle
+ *
+ * Given a TIOCE bus address, set the appropriate bit to indicate barrier
+ * attributes.
+ */
+static u64
+tioce_dma_barrier(u64 bus_addr, int on)
+{
+	u64 barrier_bit;
+
+	/* barrier not supported in M40/M40S mode */
+	if (TIOCE_M40_ADDR(bus_addr) || TIOCE_M40S_ADDR(bus_addr))
+		return bus_addr;
+
+	if (TIOCE_D64_ADDR(bus_addr))
+		barrier_bit = (1UL << 62);
+	else			/* must be m32 or d32 */
+		barrier_bit = (1UL << 30);
+
+	return (on) ? (bus_addr | barrier_bit) : (bus_addr & ~barrier_bit);
+}
+
+/**
+ * tioce_dma_unmap - release CE mapping resources
+ * @pdev: linux pci_dev representing the function
+ * @bus_addr: bus address returned by an earlier tioce_dma_map
+ * @dir: mapping direction (unused)
+ *
+ * Locate mapping resources associated with @bus_addr and release them.
+ * For mappings created using the direct modes there are no resources
+ * to release.
+ */
+void
+tioce_dma_unmap(struct pci_dev *pdev, dma_addr_t bus_addr, int dir)
+{
+	int i;
+	int port;
+	struct tioce_kernel *ce_kern;
+	struct tioce __iomem *ce_mmr;
+	unsigned long flags;
+
+	bus_addr = tioce_dma_barrier(bus_addr, 0);
+	pcidev_to_tioce(pdev, &ce_mmr, &ce_kern, &port);
+
+	/* nothing to do for D64 */
+
+	if (TIOCE_D64_ADDR(bus_addr))
+		return;
+
+	spin_lock_irqsave(&ce_kern->ce_lock, flags);
+
+	if (TIOCE_D32_ADDR(bus_addr)) {
+		if (--ce_kern->ce_port[port].dirmap_refcnt == 0) {
+			ce_kern->ce_port[port].dirmap_shadow = 0;
+			tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_dir_map[port],
+					 0);
+		}
+	} else {
+		struct tioce_dmamap *map;
+
+		list_for_each_entry(map, &ce_kern->ce_dmamap_list,
+				    ce_dmamap_list) {
+			u64 last;
+
+			last = map->pci_start + map->nbytes - 1;
+			if (bus_addr >= map->pci_start && bus_addr <= last)
+				break;
+		}
+
+		if (&map->ce_dmamap_list == &ce_kern->ce_dmamap_list) {
+			printk(KERN_WARNING
+			       "%s:  %s - no map found for bus_addr 0x%llx\n",
+			       __func__, pci_name(pdev), bus_addr);
+		} else if (--map->refcnt == 0) {
+			for (i = 0; i < map->ate_count; i++) {
+				map->ate_shadow[i] = 0;
+				tioce_mmr_storei(ce_kern, &map->ate_hw[i], 0);
+			}
+
+			list_del(&map->ce_dmamap_list);
+			kfree(map);
+		}
+	}
+
+	spin_unlock_irqrestore(&ce_kern->ce_lock, flags);
+}
+
+/**
+ * tioce_do_dma_map - map pages for PCI DMA
+ * @pdev: linux pci_dev representing the function
+ * @paddr: host physical address to map
+ * @byte_count: bytes to map
+ *
+ * This is the main wrapper for mapping host physical pages to CE PCI space.
+ * The mapping mode used is based on the device's dma_mask.
+ */
+static u64
+tioce_do_dma_map(struct pci_dev *pdev, u64 paddr, size_t byte_count,
+		 int barrier, int dma_flags)
+{
+	unsigned long flags;
+	u64 ct_addr;
+	u64 mapaddr = 0;
+	struct tioce_kernel *ce_kern;
+	struct tioce_dmamap *map;
+	int port;
+	u64 dma_mask;
+
+	dma_mask = (barrier) ? pdev->dev.coherent_dma_mask : pdev->dma_mask;
+
+	/* cards must be able to address at least 31 bits */
+	if (dma_mask < 0x7fffffffUL)
+		return 0;
+
+	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
+		ct_addr = PHYS_TO_TIODMA(paddr);
+	else
+		ct_addr = paddr;
+
+	/*
+	 * If the device can generate 64 bit addresses, create a D64 map.
+	 */
+	if (dma_mask == ~0UL) {
+		mapaddr = tioce_dma_d64(ct_addr, dma_flags);
+		if (mapaddr)
+			goto dma_map_done;
+	}
+
+	pcidev_to_tioce(pdev, NULL, &ce_kern, &port);
+
+	spin_lock_irqsave(&ce_kern->ce_lock, flags);
+
+	/*
+	 * D64 didn't work ... See if we have an existing map that covers
+	 * this address range.  Must account for devices dma_mask here since
+	 * an existing map might have been done in a mode using more pci
+	 * address bits than this device can support.
+	 */
+	list_for_each_entry(map, &ce_kern->ce_dmamap_list, ce_dmamap_list) {
+		u64 last;
+
+		last = map->ct_start + map->nbytes - 1;
+		if (ct_addr >= map->ct_start &&
+		    ct_addr + byte_count - 1 <= last &&
+		    map->pci_start <= dma_mask) {
+			map->refcnt++;
+			mapaddr = map->pci_start + (ct_addr - map->ct_start);
+			break;
+		}
+	}
+
+	/*
+	 * If we don't have a map yet, and the card can generate 40
+	 * bit addresses, try the M40/M40S modes.  Note these modes do not
+	 * support a barrier bit, so if we need a consistent map these
+	 * won't work.
+	 */
+	if (!mapaddr && !barrier && dma_mask >= 0xffffffffffUL) {
+		/*
+		 * We have two options for 40-bit mappings:  16GB "super" ATEs
+		 * and 64MB "regular" ATEs.  We'll try both if needed for a
+		 * given mapping but which one we try first depends on the
+		 * size.  For requests >64MB, prefer to use a super page with
+		 * regular as the fallback. Otherwise, try in the reverse order.
+		 */
+
+		if (byte_count > MB(64)) {
+			mapaddr = tioce_alloc_map(ce_kern, TIOCE_ATE_M40S,
+						  port, ct_addr, byte_count,
+						  dma_flags);
+			if (!mapaddr)
+				mapaddr =
+				    tioce_alloc_map(ce_kern, TIOCE_ATE_M40, -1,
+						    ct_addr, byte_count,
+						    dma_flags);
+		} else {
+			mapaddr = tioce_alloc_map(ce_kern, TIOCE_ATE_M40, -1,
+						  ct_addr, byte_count,
+						  dma_flags);
+			if (!mapaddr)
+				mapaddr =
+				    tioce_alloc_map(ce_kern, TIOCE_ATE_M40S,
+						    port, ct_addr, byte_count,
+						    dma_flags);
+		}
+	}
+
+	/*
+	 * 32-bit direct is the next mode to try
+	 */
+	if (!mapaddr && dma_mask >= 0xffffffffUL)
+		mapaddr = tioce_dma_d32(pdev, ct_addr, dma_flags);
+
+	/*
+	 * Last resort, try 32-bit ATE-based map.
+	 */
+	if (!mapaddr)
+		mapaddr =
+		    tioce_alloc_map(ce_kern, TIOCE_ATE_M32, -1, ct_addr,
+				    byte_count, dma_flags);
+
+	spin_unlock_irqrestore(&ce_kern->ce_lock, flags);
+
+dma_map_done:
+	if (mapaddr && barrier)
+		mapaddr = tioce_dma_barrier(mapaddr, 1);
+
+	return mapaddr;
+}
+
+/**
+ * tioce_dma - standard pci dma map interface
+ * @pdev: pci device requesting the map
+ * @paddr: system physical address to map into pci space
+ * @byte_count: # bytes to map
+ *
+ * Simply call tioce_do_dma_map() to create a map with the barrier bit clear
+ * in the address.
+ */
+static u64
+tioce_dma(struct pci_dev *pdev, unsigned long  paddr, size_t byte_count, int dma_flags)
+{
+	return tioce_do_dma_map(pdev, paddr, byte_count, 0, dma_flags);
+}
+
+/**
+ * tioce_dma_consistent - consistent pci dma map interface
+ * @pdev: pci device requesting the map
+ * @paddr: system physical address to map into pci space
+ * @byte_count: # bytes to map
+ *
+ * Simply call tioce_do_dma_map() to create a map with the barrier bit set
+ * in the address.
+ */
+static u64
+tioce_dma_consistent(struct pci_dev *pdev, unsigned long  paddr, size_t byte_count, int dma_flags)
+{
+	return tioce_do_dma_map(pdev, paddr, byte_count, 1, dma_flags);
+}
+
+/**
+ * tioce_error_intr_handler - SGI TIO CE error interrupt handler
+ * @irq: unused
+ * @arg: pointer to tioce_common struct for the given CE
+ *
+ * Handle a CE error interrupt.  Simply a wrapper around a SAL call which
+ * defers processing to the SGI prom.
+ */
+static irqreturn_t
+tioce_error_intr_handler(int irq, void *arg)
+{
+	struct tioce_common *soft = arg;
+	struct ia64_sal_retval ret_stuff;
+	ret_stuff.status = 0;
+	ret_stuff.v0 = 0;
+
+	SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_ERROR_INTERRUPT,
+			soft->ce_pcibus.bs_persist_segment,
+			soft->ce_pcibus.bs_persist_busnum, 0, 0, 0, 0, 0);
+
+	if (ret_stuff.v0)
+		panic("tioce_error_intr_handler:  Fatal TIOCE error");
+
+	return IRQ_HANDLED;
+}
+
+/**
+ * tioce_reserve_m32 - reserve M32 ATEs for the indicated address range
+ * @tioce_kernel: TIOCE context to reserve ATEs for
+ * @base: starting bus address to reserve
+ * @limit: last bus address to reserve
+ *
+ * If base/limit falls within the range of bus space mapped through the
+ * M32 space, reserve the resources corresponding to the range.
+ */
+static void
+tioce_reserve_m32(struct tioce_kernel *ce_kern, u64 base, u64 limit)
+{
+	int ate_index, last_ate, ps;
+	struct tioce __iomem *ce_mmr;
+
+	ce_mmr = (struct tioce __iomem *)ce_kern->ce_common->ce_pcibus.bs_base;
+	ps = ce_kern->ce_ate3240_pagesize;
+	ate_index = ATE_PAGE(base, ps);
+	last_ate = ate_index + ATE_NPAGES(base, limit-base+1, ps) - 1;
+
+	if (ate_index < 64)
+		ate_index = 64;
+
+	if (last_ate >= TIOCE_NUM_M3240_ATES)
+		last_ate = TIOCE_NUM_M3240_ATES - 1;
+
+	while (ate_index <= last_ate) {
+		u64 ate;
+
+		ate = ATE_MAKE(0xdeadbeef, ps, 0);
+		ce_kern->ce_ate3240_shadow[ate_index] = ate;
+		tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_ate3240[ate_index],
+				 ate);
+		ate_index++;
+	}
+}
+
+/**
+ * tioce_kern_init - init kernel structures related to a given TIOCE
+ * @tioce_common: ptr to a cached tioce_common struct that originated in prom
+ */
+static struct tioce_kernel *
+tioce_kern_init(struct tioce_common *tioce_common)
+{
+	int i;
+	int ps;
+	int dev;
+	u32 tmp;
+	unsigned int seg, bus;
+	struct tioce __iomem *tioce_mmr;
+	struct tioce_kernel *tioce_kern;
+
+	tioce_kern = kzalloc(sizeof(struct tioce_kernel), GFP_KERNEL);
+	if (!tioce_kern) {
+		return NULL;
+	}
+
+	tioce_kern->ce_common = tioce_common;
+	spin_lock_init(&tioce_kern->ce_lock);
+	INIT_LIST_HEAD(&tioce_kern->ce_dmamap_list);
+	tioce_common->ce_kernel_private = (u64) tioce_kern;
+
+	/*
+	 * Determine the secondary bus number of the port2 logical PPB.
+	 * This is used to decide whether a given pci device resides on
+	 * port1 or port2.  Note:  We don't have enough plumbing set up
+	 * here to use pci_read_config_xxx() so use raw_pci_read().
+	 */
+
+	seg = tioce_common->ce_pcibus.bs_persist_segment;
+	bus = tioce_common->ce_pcibus.bs_persist_busnum;
+
+	raw_pci_read(seg, bus, PCI_DEVFN(2, 0), PCI_SECONDARY_BUS, 1,&tmp);
+	tioce_kern->ce_port1_secondary = (u8) tmp;
+
+	/*
+	 * Set PMU pagesize to the largest size available, and zero out
+	 * the ATEs.
+	 */
+
+	tioce_mmr = (struct tioce __iomem *)tioce_common->ce_pcibus.bs_base;
+	tioce_mmr_clri(tioce_kern, &tioce_mmr->ce_ure_page_map,
+		       CE_URE_PAGESIZE_MASK);
+	tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_ure_page_map,
+		       CE_URE_256K_PAGESIZE);
+	ps = tioce_kern->ce_ate3240_pagesize = KB(256);
+
+	for (i = 0; i < TIOCE_NUM_M40_ATES; i++) {
+		tioce_kern->ce_ate40_shadow[i] = 0;
+		tioce_mmr_storei(tioce_kern, &tioce_mmr->ce_ure_ate40[i], 0);
+	}
+
+	for (i = 0; i < TIOCE_NUM_M3240_ATES; i++) {
+		tioce_kern->ce_ate3240_shadow[i] = 0;
+		tioce_mmr_storei(tioce_kern, &tioce_mmr->ce_ure_ate3240[i], 0);
+	}
+
+	/*
+	 * Reserve ATEs corresponding to reserved address ranges.  These
+	 * include:
+	 *
+	 *	Memory space covered by each PPB mem base/limit register
+	 * 	Memory space covered by each PPB prefetch base/limit register
+	 *
+	 * These bus ranges are for pio (downstream) traffic only, and so
+	 * cannot be used for DMA.
+	 */
+
+	for (dev = 1; dev <= 2; dev++) {
+		u64 base, limit;
+
+		/* mem base/limit */
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_MEMORY_BASE, 2, &tmp);
+		base = (u64)tmp << 16;
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_MEMORY_LIMIT, 2, &tmp);
+		limit = (u64)tmp << 16;
+		limit |= 0xfffffUL;
+
+		if (base < limit)
+			tioce_reserve_m32(tioce_kern, base, limit);
+
+		/*
+		 * prefetch mem base/limit.  The tioce ppb's have 64-bit
+		 * decoders, so read the upper portions w/o checking the
+		 * attributes.
+		 */
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_PREF_MEMORY_BASE, 2, &tmp);
+		base = ((u64)tmp & PCI_PREF_RANGE_MASK) << 16;
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_PREF_BASE_UPPER32, 4, &tmp);
+		base |= (u64)tmp << 32;
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_PREF_MEMORY_LIMIT, 2, &tmp);
+
+		limit = ((u64)tmp & PCI_PREF_RANGE_MASK) << 16;
+		limit |= 0xfffffUL;
+
+		raw_pci_read(seg, bus, PCI_DEVFN(dev, 0),
+				  PCI_PREF_LIMIT_UPPER32, 4, &tmp);
+		limit |= (u64)tmp << 32;
+
+		if ((base < limit) && TIOCE_M32_ADDR(base))
+			tioce_reserve_m32(tioce_kern, base, limit);
+	}
+
+	return tioce_kern;
+}
+
+/**
+ * tioce_force_interrupt - implement altix force_interrupt() backend for CE
+ * @sn_irq_info: sn asic irq that we need an interrupt generated for
+ *
+ * Given an sn_irq_info struct, set the proper bit in ce_adm_force_int to
+ * force a secondary interrupt to be generated.  This is to work around an
+ * asic issue where there is a small window of opportunity for a legacy device
+ * interrupt to be lost.
+ */
+static void
+tioce_force_interrupt(struct sn_irq_info *sn_irq_info)
+{
+	struct pcidev_info *pcidev_info;
+	struct tioce_common *ce_common;
+	struct tioce_kernel *ce_kern;
+	struct tioce __iomem *ce_mmr;
+	u64 force_int_val;
+
+	if (!sn_irq_info->irq_bridge)
+		return;
+
+	if (sn_irq_info->irq_bridge_type != PCIIO_ASIC_TYPE_TIOCE)
+		return;
+
+	pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	if (!pcidev_info)
+		return;
+
+	ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info;
+	ce_mmr = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base;
+	ce_kern = (struct tioce_kernel *)ce_common->ce_kernel_private;
+
+	/*
+	 * TIOCE Rev A workaround (PV 945826), force an interrupt by writing
+	 * the TIO_INTx register directly (1/26/2006)
+	 */
+	if (ce_common->ce_rev == TIOCE_REV_A) {
+		u64 int_bit_mask = (1ULL << sn_irq_info->irq_int_bit);
+		u64 status;
+
+		tioce_mmr_load(ce_kern, &ce_mmr->ce_adm_int_status, &status);
+		if (status & int_bit_mask) {
+			u64 force_irq = (1 << 8) | sn_irq_info->irq_irq;
+			u64 ctalk = sn_irq_info->irq_xtalkaddr;
+			u64 nasid, offset;
+
+			nasid = (ctalk & CTALK_NASID_MASK) >> CTALK_NASID_SHFT;
+			offset = (ctalk & CTALK_NODE_OFFSET);
+			HUB_S(TIO_IOSPACE_ADDR(nasid, offset), force_irq);
+		}
+
+		return;
+	}
+
+	/*
+	 * irq_int_bit is originally set up by prom, and holds the interrupt
+	 * bit shift (not mask) as defined by the bit definitions in the
+	 * ce_adm_int mmr.  These shifts are not the same for the
+	 * ce_adm_force_int register, so do an explicit mapping here to make
+	 * things clearer.
+	 */
+
+	switch (sn_irq_info->irq_int_bit) {
+	case CE_ADM_INT_PCIE_PORT1_DEV_A_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_A_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT1_DEV_B_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_B_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT1_DEV_C_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_C_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT1_DEV_D_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_D_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT2_DEV_A_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_A_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT2_DEV_B_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_B_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT2_DEV_C_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_C_SHFT;
+		break;
+	case CE_ADM_INT_PCIE_PORT2_DEV_D_SHFT:
+		force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_D_SHFT;
+		break;
+	default:
+		return;
+	}
+	tioce_mmr_storei(ce_kern, &ce_mmr->ce_adm_force_int, force_int_val);
+}
+
+/**
+ * tioce_target_interrupt - implement set_irq_affinity for tioce resident
+ * functions.  Note:  only applies to line interrupts, not MSI's.
+ *
+ * @sn_irq_info: SN IRQ context
+ *
+ * Given an sn_irq_info, set the associated CE device's interrupt destination
+ * register.  Since the interrupt destination registers are on a per-ce-slot
+ * basis, this will retarget line interrupts for all functions downstream of
+ * the slot.
+ */
+static void
+tioce_target_interrupt(struct sn_irq_info *sn_irq_info)
+{
+	struct pcidev_info *pcidev_info;
+	struct tioce_common *ce_common;
+	struct tioce_kernel *ce_kern;
+	struct tioce __iomem *ce_mmr;
+	int bit;
+	u64 vector;
+
+	pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo;
+	if (!pcidev_info)
+		return;
+
+	ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info;
+	ce_mmr = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base;
+	ce_kern = (struct tioce_kernel *)ce_common->ce_kernel_private;
+
+	bit = sn_irq_info->irq_int_bit;
+
+	tioce_mmr_seti(ce_kern, &ce_mmr->ce_adm_int_mask, (1UL << bit));
+	vector = (u64)sn_irq_info->irq_irq << INTR_VECTOR_SHFT;
+	vector |= sn_irq_info->irq_xtalkaddr;
+	tioce_mmr_storei(ce_kern, &ce_mmr->ce_adm_int_dest[bit], vector);
+	tioce_mmr_clri(ce_kern, &ce_mmr->ce_adm_int_mask, (1UL << bit));
+
+	tioce_force_interrupt(sn_irq_info);
+}
+
+/**
+ * tioce_bus_fixup - perform final PCI fixup for a TIO CE bus
+ * @prom_bussoft: Common prom/kernel struct representing the bus
+ *
+ * Replicates the tioce_common pointed to by @prom_bussoft in kernel
+ * space.  Allocates and initializes a kernel-only area for a given CE,
+ * and sets up an irq for handling CE error interrupts.
+ *
+ * On successful setup, returns the kernel version of tioce_common back to
+ * the caller.
+ */
+static void *
+tioce_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller)
+{
+	struct tioce_common *tioce_common;
+	struct tioce_kernel *tioce_kern;
+	struct tioce __iomem *tioce_mmr;
+
+	/*
+	 * Allocate kernel bus soft and copy from prom.
+	 */
+
+	tioce_common = kzalloc(sizeof(struct tioce_common), GFP_KERNEL);
+	if (!tioce_common)
+		return NULL;
+
+	memcpy(tioce_common, prom_bussoft, sizeof(struct tioce_common));
+	tioce_common->ce_pcibus.bs_base = (unsigned long)
+		ioremap(REGION_OFFSET(tioce_common->ce_pcibus.bs_base),
+			sizeof(struct tioce_common));
+
+	tioce_kern = tioce_kern_init(tioce_common);
+	if (tioce_kern == NULL) {
+		kfree(tioce_common);
+		return NULL;
+	}
+
+	/*
+	 * Clear out any transient errors before registering the error
+	 * interrupt handler.
+	 */
+
+	tioce_mmr = (struct tioce __iomem *)tioce_common->ce_pcibus.bs_base;
+	tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_adm_int_status_alias, ~0ULL);
+	tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_adm_error_summary_alias,
+		       ~0ULL);
+	tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_dre_comp_err_addr, 0ULL);
+
+	if (request_irq(SGI_PCIASIC_ERROR,
+			tioce_error_intr_handler,
+			IRQF_SHARED, "TIOCE error", (void *)tioce_common))
+		printk(KERN_WARNING
+		       "%s:  Unable to get irq %d.  "
+		       "Error interrupts won't be routed for "
+		       "TIOCE bus %04x:%02x\n",
+		       __func__, SGI_PCIASIC_ERROR,
+		       tioce_common->ce_pcibus.bs_persist_segment,
+		       tioce_common->ce_pcibus.bs_persist_busnum);
+
+	irq_set_handler(SGI_PCIASIC_ERROR, handle_level_irq);
+	sn_set_err_irq_affinity(SGI_PCIASIC_ERROR);
+	return tioce_common;
+}
+
+static struct sn_pcibus_provider tioce_pci_interfaces = {
+	.dma_map = tioce_dma,
+	.dma_map_consistent = tioce_dma_consistent,
+	.dma_unmap = tioce_dma_unmap,
+	.bus_fixup = tioce_bus_fixup,
+	.force_interrupt = tioce_force_interrupt,
+	.target_interrupt = tioce_target_interrupt
+};
+
+/**
+ * tioce_init_provider - init SN PCI provider ops for TIO CE
+ */
+int
+tioce_init_provider(void)
+{
+	sn_pci_provider[PCIIO_ASIC_TYPE_TIOCE] = &tioce_pci_interfaces;
+	return 0;
+}