Add Xen support

2009-03-11  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/i386/vm_param.h (VM_MIN_KERNEL_ADDRESS) [MACH_XEN]: Set to
	0x20000000.
	* i386/i386/i386asm.sym (pfn_list) [VM_MIN_KERNEL_ADDRESS ==
	LINEAR_MIN_KERNEL_ADDRESS]: Define to constant PFN_LIST.

2009-02-27  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/intel/pmap.c [MACH_HYP] (INVALIDATE_TLB): Call hyp_invlpg
	instead of flush_tlb when e - s is compile-time known to be
	PAGE_SIZE.

2008-11-27  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/configfrag.ac (enable_pae): Enable by default on the Xen
	platform.

2007-11-14  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/i386at/model_dep.c (machine_relax): New function.
	(c_boot_entry): Refuse to boot as dom0.

2007-10-17  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/i386/fpu.c [MACH_XEN]: Disable unused fpintr().

2007-08-12  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* Makefile.am (clib_routines): Add _START.
	* i386/xen/xen_boothdr: Use _START for VIRT_BASE and PADDR_OFFSET. Add
	GUEST_VERSION and XEN_ELFNOTE_FEATURES.

2007-06-13  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* i386/i386/user_ldt.h (user_ldt) [MACH_XEN]: Add alloc field.
	* i386/i386/user_ldt.c (i386_set_ldt) [MACH_XEN]: Round allocation of
	LDT to a page, set back LDT pages read/write before freeing them.
	(user_ldt_free) [MACH_XEN]: Likewise.

2007-04-18  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	* device/ds_routines.c [MACH_HYP]: Add hypervisor block and net devices.

2007-02-19  Thomas Schwinge  <tschwinge@gnu.org>

	* i386/xen/Makefrag.am [PLATFORM_xen] (gnumach_LINKFLAGS): Define.

	* Makefrag.am: Include `xen/Makefrag.am'.
	* configure.ac: Include `xen/configfrag.ac'.
	(--enable-platform): Support the `xen' platform.
	* i386/configfrag.ac: Likewise.
	* i386/Makefrag.am [PLATFORM_xen]: Include `i386/xen/Makefrag.am'.

2007-02-19  Samuel Thibault  <samuel.thibault@ens-lyon.org>
	    Thomas Schwinge  <tschwinge@gnu.org>

	* i386/xen/Makefrag.am: New file.
	* xen/Makefrag.am: Likewise.
	* xen/configfrag.ac: Likewise.

2007-02-11 (and later dates)  Samuel Thibault  <samuel.thibault@ens-lyon.org>

	Xen support
	* Makefile.am (clib_routines): Add _start.
	* Makefrag.am (include_mach_HEADERS): Add include/mach/xen.h.
	* device/cons.c (cnputc): Call hyp_console_write.
	* i386/Makefrag.am (libkernel_a_SOURCES): Move non-Xen source to
	[PLATFORM_at].
	* i386/i386/debug_trace.S: Include <i386/xen.h>
	* i386/i386/fpu.c [MACH_HYP] (init_fpu): Call set_ts() and clear_ts(),
	do not enable CR0_EM.
	* i386/i386/gdt.c: Include <mach/xen.h> and <intel/pmap.h>.
	[MACH_XEN]: Make gdt array extern.
	[MACH_XEN] (gdt_init): Register gdt with hypervisor.  Request 4gb
	segments assist.  Shift la_shift.
	[MACH_PSEUDO_PHYS] (gdt_init): Shift pfn_list.
	* i386/i386/gdt.h [MACH_XEN]: Don't define KERNEL_LDT and LINEAR_DS.
	* i386/i386/i386asm.sym: Include <i386/xen.h>.
	[MACH_XEN]: Remove KERNEL_LDT, Add shared_info's CPU_CLI, CPU_PENDING,
	CPU_PENDING_SEL, PENDING, EVTMASK and CR2.
	* i386/i386/idt.c [MACH_HYP] (idt_init): Register trap table with
	hypervisor.
	* i386/i386/idt_inittab.S: Include <i386/i386asm.h>.
	[MACH_XEN]: Set IDT_ENTRY() for hypervisor. Set trap table terminator.
	* i386/i386/ktss.c [MACH_XEN] (ktss_init): Request exception task switch
	from hypervisor.
	* i386/i386/ldt.c: Include <mach/xen.h> and <intel/pmap.h>
	[MACH_XEN]: Make ldt array extern.
	[MACH_XEN] (ldt_init): Set ldt readwrite.
	[MACH_HYP] (ldt_init): Register ldt with hypervisor.
	* i386/i386/locore.S: Include <i386/xen.h>. Handle KERNEL_RING == 1
	case.
	[MACH_XEN]: Read hyp_shared_info's CR2 instead of %cr2.
	[MACH_PSEUDO_PHYS]: Add mfn_to_pfn computation.
	[MACH_HYP]: Drop Cyrix I/O-based detection.  Read cr3 instead of %cr3.
	Make hypervisor call for pte invalidation.
	* i386/i386/mp_desc.c: Include <mach/xen.h>.
	[MACH_HYP] (mp_desc_init): Panic.
	* i386/i386/pcb.c: Include <mach/xen.h>.
	[MACH_XEN] (switch_ktss): Request stack switch from hypervisor.
	[MACH_HYP] (switch_ktss): Request ldt and gdt switch from hypervisor.
	* i386/i386/phys.c: Include <mach/xen.h>
	[MACH_PSEUDO_PHYS] (kvtophys): Do page translation.
	* i386/i386/proc_reg.h [MACH_HYP] (cr3): New declaration.
	(set_cr3, get_cr3, set_ts, clear_ts): Implement macros.
	* i386/i386/seg.h [MACH_HYP]: Define KERNEL_RING macro.  Include
	<mach/xen.h>
	[MACH_XEN] (fill_descriptor): Register descriptor with hypervisor.
	* i386/i386/spl.S: Include <i386/xen.h> and <i386/i386/asm.h>
	[MACH_XEN] (pic_mask): #define to int_mask.
	[MACH_XEN] (SETMASK): Implement.
	* i386/i386/vm_param.h [MACH_XEN] (HYP_VIRT_START): New macro.
	[MACH_XEN]: Set VM_MAX_KERNEL_ADDRESS to HYP_VIRT_START-
	LINEAR_MIN_KERNEL_ADDRESS + VM_MIN_KERNEL_ADDRESS.  Increase
	KERNEL_STACK_SIZE and INTSTACK_SIZE to 4 pages.
	* i386/i386at/conf.c [MACH_HYP]: Remove hardware devices, add hypervisor
	console device.
	* i386/i386at/cons_conf.c [MACH_HYP]: Add hypervisor console device.
	* i386/i386at/model_dep.c: Include <sys/types.h>, <mach/xen.h>.
	[MACH_XEN] Include <xen/console.h>, <xen/store.h>, <xen/evt.h>,
	<xen/xen.h>.
	[MACH_PSEUDO_PHYS]: New boot_info, mfn_list, pfn_list variables.
	[MACH_XEN]: New la_shift variable.
	[MACH_HYP] (avail_next, mem_size_init): Drop BIOS skipping mecanism.
	[MACH_HYP] (machine_init): Call hyp_init(), drop hardware
	initialization.
	[MACH_HYP] (machine_idle): Call hyp_idle().
	[MACH_HYP] (halt_cpu): Call hyp_halt().
	[MACH_HYP] (halt_all_cpus): Call hyp_reboot() or hyp_halt().
	[MACH_HYP] (i386at_init): Initialize with hypervisor.
	[MACH_XEN] (c_boot_entry): Add Xen-specific initialization.
	[MACH_HYP] (init_alloc_aligned, pmap_valid_page): Drop zones skipping
	mecanism.
	* i386/intel/pmap.c: Include <mach/xen.h>.
	[MACH_PSEUDO_PHYS] (WRITE_PTE): Do page translation.
	[MACH_HYP] (INVALIDATE_TLB): Request invalidation from hypervisor.
	[MACH_XEN] (pmap_map_bd, pmap_create, pmap_destroy, pmap_remove_range)
	(pmap_page_protect, pmap_protect, pmap_enter, pmap_change_wiring)
	(pmap_attribute_clear, pmap_unmap_page_zero, pmap_collect): Request MMU
	update from hypervisor.
	[MACH_XEN] (pmap_bootstrap): Request pagetable initialization from
	hypervisor.
	[MACH_XEN] (pmap_set_page_readwrite, pmap_set_page_readonly)
	(pmap_set_page_readonly_init, pmap_clear_bootstrap_pagetable)
	(pmap_map_mfn): New functions.
	* i386/intel/pmap.h [MACH_XEN] (INTEL_PTE_GLOBAL): Disable global page
	support.
	[MACH_PSEUDO_PHYS] (pte_to_pa): Do page translation.
	[MACH_XEN] (pmap_set_page_readwrite, pmap_set_page_readonly)
	(pmap_set_page_readonly_init, pmap_clear_bootstrap_pagetable)
	(pmap_map_mfn): Declare functions.
	* i386/i386/xen.h: New file.
	* i386/xen/xen.c: New file.
	* i386/xen/xen_boothdr.S: New file.
	* i386/xen/xen_locore.S: New file.
	* include/mach/xen.h: New file.
	* kern/bootstrap.c [MACH_XEN] (boot_info): Declare variable.
	[MACH_XEN] (bootstrap_create): Rebase multiboot header.
	* kern/debug.c: Include <mach/xen.h>.
	[MACH_HYP] (panic): Call hyp_crash() without delay.
	* linux/dev/include/asm-i386/segment.h [MACH_HYP] (KERNEL_CS)
	(KERNEL_DS): Use ring 1.
	* xen/block.c: New file.
	* xen/block.h: Likewise.
	* xen/console.c: Likewise.
	* xen/console.h: Likewise.
	* xen/evt.c: Likewise.
	* xen/evt.h: Likewise.
	* xen/grant.c: Likewise.
	* xen/grant.h: Likewise.
	* xen/net.c: Likewise.
	* xen/net.h: Likewise.
	* xen/ring.c: Likewise.
	* xen/ring.h: Likewise.
	* xen/store.c: Likewise.
	* xen/store.h: Likewise.
	* xen/time.c: Likewise.
	* xen/time.h: Likewise.
	* xen/xen.c: Likewise.
	* xen/xen.h: Likewise.
	* xen/public/COPYING: Import file from Xen.
	* xen/public/callback.h: Likewise.
	* xen/public/dom0_ops.h: Likewise.
	* xen/public/domctl.h: Likewise.
	* xen/public/elfnote.h: Likewise.
	* xen/public/elfstructs.h: Likewise.
	* xen/public/event_channel.h: Likewise.
	* xen/public/features.h: Likewise.
	* xen/public/grant_table.h: Likewise.
	* xen/public/kexec.h: Likewise.
	* xen/public/libelf.h: Likewise.
	* xen/public/memory.h: Likewise.
	* xen/public/nmi.h: Likewise.
	* xen/public/physdev.h: Likewise.
	* xen/public/platform.h: Likewise.
	* xen/public/sched.h: Likewise.
	* xen/public/sysctl.h: Likewise.
	* xen/public/trace.h: Likewise.
	* xen/public/vcpu.h: Likewise.
	* xen/public/version.h: Likewise.
	* xen/public/xen-compat.h: Likewise.
	* xen/public/xen.h: Likewise.
	* xen/public/xencomm.h: Likewise.
	* xen/public/xenoprof.h: Likewise.
	* xen/public/arch-x86/xen-mca.h: Likewise.
	* xen/public/arch-x86/xen-x86_32.h: Likewise.
	* xen/public/arch-x86/xen-x86_64.h: Likewise.
	* xen/public/arch-x86/xen.h: Likewise.
	* xen/public/arch-x86_32.h: Likewise.
	* xen/public/arch-x86_64.h: Likewise.
	* xen/public/io/blkif.h: Likewise.
	* xen/public/io/console.h: Likewise.
	* xen/public/io/fbif.h: Likewise.
	* xen/public/io/fsif.h: Likewise.
	* xen/public/io/kbdif.h: Likewise.
	* xen/public/io/netif.h: Likewise.
	* xen/public/io/pciif.h: Likewise.
	* xen/public/io/protocols.h: Likewise.
	* xen/public/io/ring.h: Likewise.
	* xen/public/io/tpmif.h: Likewise.
	* xen/public/io/xenbus.h: Likewise.
	* xen/public/io/xs_wire.h: Likewise.

3 files changed, 195 insertions, 14 deletions

diff --git a/i386/i386at/conf.c b/i386/i386at/conf.c
index 23c2a6f..f5ab36c 100644
--- a/i386/i386at/conf.c
+++ b/i386/i386at/conf.c
@@ -34,6 +34,7 @@ extern int	timeopen(), timeclose();
 extern vm_offset_t timemmap();
 #define	timename		"time"
 
+#ifndef	MACH_HYP
 extern int	kdopen(), kdclose(), kdread(), kdwrite();
 extern int	kdgetstat(), kdsetstat(), kdportdeath();
 extern vm_offset_t kdmmap();
@@ -50,17 +51,26 @@ extern int	lpropen(), lprclose(), lprread(), lprwrite();
 extern int	lprgetstat(), lprsetstat(), lprportdeath();
 #define	lprname			"lpr"
 #endif	/* NLPR > 0 */
+#endif	/* MACH_HYP */
 
 extern int	kbdopen(), kbdclose(), kbdread();
 extern int	kbdgetstat(), kbdsetstat();
 #define	kbdname			"kbd"
 
+#ifndef	MACH_HYP
 extern int	mouseopen(), mouseclose(), mouseread(), mousegetstat();
 #define	mousename		"mouse"
+#endif	/* MACH_HYP */
 
 extern int	kmsgopen(), kmsgclose(), kmsgread(), kmsggetstat();
 #define kmsgname		"kmsg"
 
+#ifdef	MACH_HYP
+extern int	hypcnopen(), hypcnclose(), hypcnread(), hypcnwrite();
+extern int	hypcngetstat(), hypcnsetstat(), hypcnportdeath();
+#define hypcnname		"hyp"
+#endif	/* MACH_HYP */
+
 /*
  * List of devices - console must be at slot 0
  */
@@ -79,16 +89,19 @@ struct dev_ops	dev_name_list[] =
 	  nodev,	nulldev,	nulldev,	0,
 	  nodev },
 
+#ifndef	MACH_HYP
 	{ kdname,	kdopen,		kdclose,	kdread,
 	  kdwrite,	kdgetstat,	kdsetstat,	kdmmap,
 	  nodev,	nulldev,	kdportdeath,	0,
 	  nodev },
+#endif	/* MACH_HYP */
 
 	{ timename,	timeopen,	timeclose,	nulldev,
 	  nulldev,	nulldev,	nulldev,	timemmap,
 	  nodev,	nulldev,	nulldev,	0,
 	  nodev },
 
+#ifndef	MACH_HYP
 #if	NCOM > 0
 	{ comname,	comopen,	comclose,	comread,
 	  comwrite,	comgetstat,	comsetstat,	nomap,
@@ -107,6 +120,7 @@ struct dev_ops	dev_name_list[] =
 	  nodev,	mousegetstat,	nulldev,	nomap,
 	  nodev,	nulldev,	nulldev,	0,
 	  nodev },
+#endif	/* MACH_HYP */
 
 	{ kbdname,	kbdopen,	kbdclose,	kbdread,
 	  nodev,	kbdgetstat,	kbdsetstat,	nomap,
@@ -120,6 +134,13 @@ struct dev_ops	dev_name_list[] =
           nodev },
 #endif
 
+#ifdef	MACH_HYP
+	{ hypcnname,	hypcnopen,	hypcnclose,	hypcnread,
+	  hypcnwrite,	hypcngetstat,	hypcnsetstat,	nomap,
+	  nodev,	nulldev,	hypcnportdeath,	0,
+	  nodev },
+#endif	/* MACH_HYP */
+
 };
 int	dev_name_count = sizeof(dev_name_list)/sizeof(dev_name_list[0]);
 
diff --git a/i386/i386at/cons_conf.c b/i386/i386at/cons_conf.c
index 8784ed9..ea8ccb5 100644
--- a/i386/i386at/cons_conf.c
+++ b/i386/i386at/cons_conf.c
@@ -30,19 +30,27 @@
 #include <sys/types.h>
 #include <device/cons.h>
 
+#ifdef	MACH_HYP
+extern	int hypcnprobe(), hypcninit(), hypcngetc(), hypcnputc();
+#else	/* MACH_HYP */
 extern	int kdcnprobe(), kdcninit(), kdcngetc(), kdcnputc();
 #if NCOM > 0 && RCLINE >= 0
 extern	int comcnprobe(), comcninit(), comcngetc(), comcnputc();
 #endif
+#endif	/* MACH_HYP */
 
 /*
  * The rest of the consdev fields are filled in by the respective
  * cnprobe routine.
  */
 struct	consdev constab[] = {
+#ifdef	MACH_HYP
+	{"hyp",	hypcnprobe,	hypcninit,	hypcngetc,	hypcnputc},
+#else	/* MACH_HYP */
 	{"kd",	kdcnprobe,	kdcninit,	kdcngetc,	kdcnputc},
 #if NCOM > 0 && RCLINE >= 0 && 1
 	{"com",	comcnprobe,	comcninit,	comcngetc,	comcnputc},
 #endif
+#endif	/* MACH_HYP */
 	{0}
 };
diff --git a/i386/i386at/model_dep.c b/i386/i386at/model_dep.c
index 3ebe2e6..61605a1 100644
--- a/i386/i386at/model_dep.c
+++ b/i386/i386at/model_dep.c
@@ -40,6 +40,7 @@
 #include <mach/vm_prot.h>
 #include <mach/machine.h>
 #include <mach/machine/multiboot.h>
+#include <mach/xen.h>
 
 #include <i386/vm_param.h>
 #include <kern/assert.h>
@@ -48,6 +49,7 @@
 #include <kern/mach_clock.h>
 #include <kern/printf.h>
 #include <sys/time.h>
+#include <sys/types.h>
 #include <vm/vm_page.h>
 #include <i386/fpu.h>
 #include <i386/gdt.h>
@@ -65,6 +67,12 @@
 #include <i386at/int_init.h>
 #include <i386at/kd.h>
 #include <i386at/rtc.h>
+#ifdef	MACH_XEN
+#include <xen/console.h>
+#include <xen/store.h>
+#include <xen/evt.h>
+#include <xen/xen.h>
+#endif	/* MACH_XEN */
 
 /* Location of the kernel's symbol table.
    Both of these are 0 if none is available.  */
@@ -81,7 +89,20 @@ vm_offset_t phys_first_addr = 0;
 vm_offset_t phys_last_addr;
 
 /* A copy of the multiboot info structure passed by the boot loader.  */
+#ifdef MACH_XEN
+struct start_info boot_info;
+#ifdef MACH_PSEUDO_PHYS
+unsigned long *mfn_list;
+#if VM_MIN_KERNEL_ADDRESS != LINEAR_MIN_KERNEL_ADDRESS
+unsigned long *pfn_list = (void*) PFN_LIST;
+#endif
+#endif	/* MACH_PSEUDO_PHYS */
+#if VM_MIN_KERNEL_ADDRESS != LINEAR_MIN_KERNEL_ADDRESS
+unsigned long la_shift = VM_MIN_KERNEL_ADDRESS;
+#endif
+#else	/* MACH_XEN */
 struct multiboot_info boot_info;
+#endif	/* MACH_XEN */
 
 /* Command line supplied to kernel.  */
 char *kernel_cmdline = "";
@@ -90,7 +111,11 @@ char *kernel_cmdline = "";
    it gets bumped up through physical memory
    that exists and is not occupied by boot gunk.
    It is not necessarily page-aligned.  */
-static vm_offset_t avail_next = 0x1000; /* XX end of BIOS data area */
+static vm_offset_t avail_next
+#ifndef MACH_HYP
+	= 0x1000 /* XX end of BIOS data area */
+#endif	/* MACH_HYP */
+	;
 
 /* Possibly overestimated amount of available memory
    still remaining to be handed to the VM system.  */
@@ -135,6 +160,9 @@ void machine_init(void)
 	 */
 	init_fpu();
 
+#ifdef MACH_HYP
+	hyp_init();
+#else	/* MACH_HYP */
 #ifdef LINUX_DEV
 	/*
 	 * Initialize Linux drivers.
@@ -146,16 +174,19 @@ void machine_init(void)
 	 * Find the devices
 	 */
 	probeio();
+#endif	/* MACH_HYP */
 
 	/*
 	 * Get the time
 	 */
 	inittodr();
 
+#ifndef MACH_HYP
 	/*
 	 * Tell the BIOS not to clear and test memory.
 	 */
 	*(unsigned short *)phystokv(0x472) = 0x1234;
+#endif	/* MACH_HYP */
 
 	/*
 	 * Unmap page 0 to trap NULL references.
@@ -166,8 +197,17 @@ void machine_init(void)
 /* Conserve power on processor CPU.  */
 void machine_idle (int cpu)
 {
+#ifdef	MACH_HYP
+  hyp_idle();
+#else	/* MACH_HYP */
   assert (cpu == cpu_number ());
   asm volatile ("hlt" : : : "memory");
+#endif	/* MACH_HYP */
+}
+
+void machine_relax ()
+{
+	asm volatile ("rep; nop" : : : "memory");
 }
 
 /*
@@ -175,9 +215,13 @@ void machine_idle (int cpu)
  */
 void halt_cpu(void)
 {
+#ifdef	MACH_HYP
+	hyp_halt();
+#else	/* MACH_HYP */
 	asm volatile("cli");
 	while (TRUE)
 	  machine_idle (cpu_number ());
+#endif	/* MACH_HYP */
 }
 
 /*
@@ -187,10 +231,16 @@ void halt_all_cpus(reboot)
 	boolean_t	reboot;
 {
 	if (reboot) {
+#ifdef	MACH_HYP
+	    hyp_reboot();
+#endif	/* MACH_HYP */
 	    kdreboot();
 	}
 	else {
 	    rebootflag = 1;
+#ifdef	MACH_HYP
+	    hyp_halt();
+#endif	/* MACH_HYP */
 	    printf("In tight loop: hit ctl-alt-del to reboot\n");
 	    (void) spl0();
 	}
@@ -215,22 +265,26 @@ void db_reset_cpu(void)
 void
 mem_size_init(void)
 {
-	vm_size_t phys_last_kb;
-
 	/* Physical memory on all PCs starts at physical address 0.
 	   XX make it a constant.  */
 	phys_first_addr = 0;
 
-	phys_last_kb = 0x400 + boot_info.mem_upper;
+#ifdef MACH_HYP
+	if (boot_info.nr_pages >= 0x100000) {
+		printf("Truncating memory size to 4GiB\n");
+		phys_last_addr = 0xffffffffU;
+	} else
+		phys_last_addr = boot_info.nr_pages * 0x1000;
+#else	/* MACH_HYP */
+	/* TODO: support mmap */
+	vm_size_t phys_last_kb = 0x400 + boot_info.mem_upper;
 	/* Avoid 4GiB overflow.  */
 	if (phys_last_kb < 0x400 || phys_last_kb >= 0x400000) {
 		printf("Truncating memory size to 4GiB\n");
-		phys_last_kb = 0x400000 - 1;
-	}
-
-	/* TODO: support mmap */
-
-	phys_last_addr = phys_last_kb * 0x400;
+		phys_last_addr = 0xffffffffU;
+	} else
+		phys_last_addr = phys_last_kb * 0x400;
+#endif	/* MACH_HYP */
 
 	printf("AT386 boot: physical memory from 0x%x to 0x%x\n",
 	       phys_first_addr, phys_last_addr);
@@ -240,14 +294,20 @@ mem_size_init(void)
 	if (phys_last_addr > ((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / 6) * 5) {
 		phys_last_addr = ((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / 6) * 5;
 		printf("Truncating memory size to %dMiB\n", (phys_last_addr - phys_first_addr) / (1024 * 1024));
+		/* TODO Xen: free lost memory */
 	}
 
 	phys_first_addr = round_page(phys_first_addr);
 	phys_last_addr = trunc_page(phys_last_addr);
 
+#ifdef MACH_HYP
+	/* Memory is just contiguous */
+	avail_remaining = phys_last_addr;
+#else	/* MACH_HYP */
 	avail_remaining
 	  = phys_last_addr - (0x100000 - (boot_info.mem_lower * 0x400)
 			      - 0x1000);
+#endif	/* MACH_HYP */
 }
 
 /*
@@ -263,13 +323,20 @@ i386at_init(void)
 	/*
 	 * Initialize the PIC prior to any possible call to an spl.
 	 */
+#ifndef	MACH_HYP
 	picinit();
+#else	/* MACH_HYP */
+	hyp_intrinit();
+#endif	/* MACH_HYP */
 
 	/*
 	 * Find memory size parameters.
 	 */
 	mem_size_init();
 
+#ifdef MACH_XEN
+	kernel_cmdline = (char*) boot_info.cmd_line;
+#else	/* MACH_XEN */
 	/* Copy content pointed by boot_info before losing access to it when it
 	 * is too far in physical memory.  */
 	if (boot_info.flags & MULTIBOOT_CMDLINE) {
@@ -304,6 +371,7 @@ i386at_init(void)
 			m[i].string = addr;
 		}
 	}
+#endif	/* MACH_XEN */
 
 	/*
 	 *	Initialize kernel physical map, mapping the
@@ -325,19 +393,42 @@ i386at_init(void)
 	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS)] =
 		kernel_page_dir[lin2pdenum(LINEAR_MIN_KERNEL_ADDRESS)];
 #if PAE
+	/* PAE page tables are 2MB only */
 	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 1] =
 		kernel_page_dir[lin2pdenum(LINEAR_MIN_KERNEL_ADDRESS) + 1];
+	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 2] =
+		kernel_page_dir[lin2pdenum(LINEAR_MIN_KERNEL_ADDRESS) + 2];
+#endif	/* PAE */
+#ifdef	MACH_XEN
+	{
+		int i;
+		for (i = 0; i < PDPNUM; i++)
+			pmap_set_page_readonly_init((void*) kernel_page_dir + i * INTEL_PGBYTES);
+#if PAE
+		pmap_set_page_readonly_init(kernel_pmap->pdpbase);
+#endif	/* PAE */
+	}
+#endif	/* MACH_XEN */
+#if PAE
 	set_cr3((unsigned)_kvtophys(kernel_pmap->pdpbase));
+#ifndef	MACH_HYP
 	if (!CPU_HAS_FEATURE(CPU_FEATURE_PAE))
 		panic("CPU doesn't have support for PAE.");
 	set_cr4(get_cr4() | CR4_PAE);
+#endif	/* MACH_HYP */
 #else
 	set_cr3((unsigned)_kvtophys(kernel_page_dir));
 #endif	/* PAE */
+#ifndef	MACH_HYP
 	if (CPU_HAS_FEATURE(CPU_FEATURE_PGE))
 		set_cr4(get_cr4() | CR4_PGE);
+	/* already set by Hypervisor */
 	set_cr0(get_cr0() | CR0_PG | CR0_WP);
+#endif	/* MACH_HYP */
 	flush_instr_queue();
+#ifdef	MACH_XEN
+	pmap_clear_bootstrap_pagetable((void *)boot_info.pt_base);
+#endif	/* MACH_XEN */
 
 	/* Interrupt stacks are allocated in physical memory,
 	   while kernel stacks are allocated in kernel virtual memory,
@@ -349,18 +440,47 @@ i386at_init(void)
 	 */
 	gdt_init();
 	idt_init();
+#ifndef	MACH_HYP
 	int_init();
+#endif	/* MACH_HYP */
 	ldt_init();
 	ktss_init();
 
 	/* Get rid of the temporary direct mapping and flush it out of the TLB.  */
+#ifdef	MACH_XEN
+#ifdef	MACH_PSEUDO_PHYS
+	if (!hyp_mmu_update_pte(kv_to_ma(&kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS)]), 0))
+#else	/* MACH_PSEUDO_PHYS */
+	if (hyp_do_update_va_mapping(VM_MIN_KERNEL_ADDRESS, 0, UVMF_INVLPG | UVMF_ALL))
+#endif	/* MACH_PSEUDO_PHYS */
+		printf("couldn't unmap frame 0\n");
+#if PAE
+#ifdef	MACH_PSEUDO_PHYS
+	if (!hyp_mmu_update_pte(kv_to_ma(&kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 1]), 0))
+#else	/* MACH_PSEUDO_PHYS */
+	if (hyp_do_update_va_mapping(VM_MIN_KERNEL_ADDRESS + INTEL_PGBYTES, 0, UVMF_INVLPG | UVMF_ALL))
+#endif	/* MACH_PSEUDO_PHYS */
+		printf("couldn't unmap frame 1\n");
+#ifdef	MACH_PSEUDO_PHYS
+	if (!hyp_mmu_update_pte(kv_to_ma(&kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 2]), 0))
+#else	/* MACH_PSEUDO_PHYS */
+	if (hyp_do_update_va_mapping(VM_MIN_KERNEL_ADDRESS + 2*INTEL_PGBYTES, 0, UVMF_INVLPG | UVMF_ALL))
+#endif	/* MACH_PSEUDO_PHYS */
+		printf("couldn't unmap frame 2\n");
+#endif	/* PAE */
+	hyp_free_page(0, (void*) VM_MIN_KERNEL_ADDRESS);
+#else	/* MACH_XEN */
 	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS)] = 0;
 #if PAE
 	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 1] = 0;
+	kernel_page_dir[lin2pdenum(VM_MIN_KERNEL_ADDRESS) + 2] = 0;
 #endif	/* PAE */
+#endif	/* MACH_XEN */
 	flush_tlb();
 
-
+#ifdef	MACH_XEN
+	hyp_p2m_init();
+#endif	/* MACH_XEN */
 
 	/* XXX We'll just use the initialization stack we're already running on
 	   as the interrupt stack for now.  Later this will have to change,
@@ -384,6 +504,15 @@ void c_boot_entry(vm_offset_t bi)
 	printf(version);
 	printf("\n");
 
+#ifdef MACH_XEN
+	printf("Running on %s.\n", boot_info.magic);
+	if (boot_info.flags & SIF_PRIVILEGED)
+		panic("Mach can't run as dom0.");
+#ifdef MACH_PSEUDO_PHYS
+	mfn_list = (void*)boot_info.mfn_list;
+#endif
+#else	/* MACH_XEN */
+
 #if	MACH_KDB
 	/*
 	 * Locate the kernel's symbol table, if the boot loader provided it.
@@ -405,6 +534,7 @@ void c_boot_entry(vm_offset_t bi)
 		       symtab_size, strtab_size);
 	}
 #endif	/* MACH_KDB */
+#endif	/* MACH_XEN */
 
 	cpu_type = discover_x86_cpu_type ();
 
@@ -525,6 +655,12 @@ boolean_t
 init_alloc_aligned(vm_size_t size, vm_offset_t *addrp)
 {
 	vm_offset_t addr;
+
+#ifdef MACH_HYP
+	/* There is none */
+	if (!avail_next)
+		avail_next = _kvtophys(boot_info.pt_base) + (boot_info.nr_pt_frames + 3) * 0x1000;
+#else	/* MACH_HYP */
 	extern char start[], end[];
 	int i;
 	static int wrapped = 0;
@@ -543,11 +679,14 @@ init_alloc_aligned(vm_size_t size, vm_offset_t *addrp)
 		: 0;
 
 	retry:
+#endif	/* MACH_HYP */
 
 	/* Page-align the start address.  */
 	avail_next = round_page(avail_next);
 
+#ifndef MACH_HYP
 	/* Start with memory above 16MB, reserving the low memory for later. */
+	/* Don't care on Xen */
 	if (!wrapped && phys_last_addr > 16 * 1024*1024)
 	  {
 	    if (avail_next < 16 * 1024*1024)
@@ -563,9 +702,15 @@ init_alloc_aligned(vm_size_t size, vm_offset_t *addrp)
 		wrapped = 1;
 	      }
 	  }
+#endif	/* MACH_HYP */
 
 	/* Check if we have reached the end of memory.  */
-        if (avail_next == (wrapped ? 16 * 1024*1024 : phys_last_addr))
+        if (avail_next == 
+		(
+#ifndef MACH_HYP
+		wrapped ? 16 * 1024*1024 : 
+#endif	/* MACH_HYP */
+		phys_last_addr))
 		return FALSE;
 
 	/* Tentatively assign the current location to the caller.  */
@@ -575,6 +720,7 @@ init_alloc_aligned(vm_size_t size, vm_offset_t *addrp)
 	   and see where that puts us.  */
 	avail_next += size;
 
+#ifndef MACH_HYP
 	/* Skip past the I/O and ROM area.  */
 	if ((avail_next > (boot_info.mem_lower * 0x400)) && (addr < 0x100000))
 	{
@@ -620,6 +766,7 @@ init_alloc_aligned(vm_size_t size, vm_offset_t *addrp)
 			/* XXX string */
 		}
 	}
+#endif	/* MACH_HYP */
 
 	avail_remaining -= size;
 
@@ -649,6 +796,11 @@ boolean_t pmap_valid_page(x)
 	vm_offset_t x;
 {
 	/* XXX is this OK?  What does it matter for?  */
-	return (((phys_first_addr <= x) && (x < phys_last_addr)) &&
-		!(((boot_info.mem_lower * 1024) <= x) && (x < 1024*1024)));
+	return (((phys_first_addr <= x) && (x < phys_last_addr))
+#ifndef MACH_HYP
+		&& !(
+		((boot_info.mem_lower * 1024) <= x) && 
+		(x < 1024*1024))
+#endif	/* MACH_HYP */
+		);
 }