hot reboot for nix.  big thanks to cinap for diagnosing
my issue with trying to execute the header.  D'oh!

Reference: /n/atom/patch/applied/nixreboot
Date: Sat Mar 22 02:02:38 CET 2014
Signed-off-by: quanstro@quanstro.net

--- /sys/src/nix/k10/mkfile	Sat Mar 22 02:02:03 2014
+++ /sys/src/nix/k10/mkfile	Sat Mar 22 02:02:04 2014
@@ -43,7 +43,7 @@
 fpu.$O:		amd64.h
 fpu.$O:		/$objtype/include/ureg.h
 ioapic.$O:	apic.h io.h
-main.$O:	init.h
+main.$O:	init.h reboot.h
 memory.$O:	amd64.h
 mmu.$O:		amd64.h
 mp.$O:		apic.h
@@ -80,6 +80,8 @@
 devacpi.$O: acpi.h
 physalloc.$O: acpi.h
 
+main.$O:	reboot.h
+
 sipi.h:D:		l64sipi.$O
 	$LD -o l64sipi.out -T0xfffffffff0003000 -R4 -l -s $prereq
 	{echo 'uchar sipihandler[]={'
@@ -88,3 +90,11 @@
 	     -e '1,2d' -e '3s/^ .. .. .. .. .. .. .. ..//'	\
 	     -e 's/ ([0-9a-f][0-9a-f])/0x\1,/g'
 	 echo '};'} > $target
+
+reboot.h:D:	rebootcode.s  amd64l.h
+	$AS rebootcode.s
+	$LD -l -s -T0x11000 -R1 -o reboot.out rebootcode.$O
+	{echo 'uchar rebootcode[]={'
+	dd -bs 1 -iseek 40 -if reboot.out | xd -1x |
+		sed -e 's/^[0-9a-f]+ //' -e 's/ ([0-9a-f][0-9a-f])/0x\1,/g'
+	 echo '};'} > reboot.h
--- /sys/src/nix/k10/main.c	Sat Mar 22 02:02:06 2014
+++ /sys/src/nix/k10/main.c	Sat Mar 22 02:02:07 2014
@@ -8,6 +8,9 @@
 #include "io.h"
 #include "apic.h"
 
+#include "amd64.h"
+#include "reboot.h"
+
 Sys*	sys;
 char	dbgflg[256];
 
@@ -154,11 +157,6 @@
 	mallocinit();
 	archpciinit();
 
-	/*
-	 * This is necessary with GRUB and QEMU. Without it an interrupt can occur
-	 * at a weird vector, because the vector base is likely different, causing
-	 * havoc. Do it before any APIC initialisation.
-	 */
 	i8259init(32);
 	if(getconf("*maxmach") != nil)
 		maxmach = atoi(getconf("*maxmach"));
@@ -337,12 +335,14 @@
 	i = (int)(uintptr)v;
 	procwired(up, i);
 	splhi();
+	lapicpri(0xff);
 	m->online = 0;
 	adec(&active.nonline);
 	ndnr();
 }
 
-#include "amd64.h"
+#define REBOOTADDR (0x11000)
+
 void
 reboot(void *entry, void *code, usize size)
 {
@@ -350,7 +350,7 @@
 	uintptr a;
 	void (*f)(uintmem, uintmem, usize);
 
-//	writeconf();
+	writeconf();
 	procwired(up, 0);
 
 	for(i = 1; i < active.nonline; i++)
@@ -361,34 +361,33 @@
 	synccons();
 	devtabshutdown();
 	pcireset();
-
 	splhi();
 	lapicpri(0xff);
-	outb(0x21, 0xff);		/* 8259 pic 0: mask irqs */
-	outb(0xa1, 0xff);		/* 8259 pic 1: mask irqs */
-	up = nil;
-	m->proc = nil;
-		/* flush mmu */
 
-//	m->mmuptp[0] = 0 | PtePS | PteP | PteRW;
-//	m->pml4[0] = m->pml4[PTLX(KZERO, 3)];
-	putcr3(m->pml4->pa);
+	if(up != nil){
+		m->proc = nil;
+		mmurelease(up);
+		up = nil;
+	}
 
-//	f = (void*)REBOOTADDR;
-//	memmove(f, rebootcode, sizeof(rebootcode));
-	f = (void*)0;
+	mmumap(0, 0, 6*1024*1024, PteP|PteRW);
+	putcr3(m->pml4->pa);
 
-	(*f)(PADDR(entry), PADDR(code), size);
+	f = UINT2PTR(REBOOTADDR);
+	memmove(f, rebootcode, sizeof(rebootcode));
+	print("warp32...\n");
+	f(PTR2UINT(entry), PADDR(code), size);
 }
 
 void
 exit(int ispanic)
 {
-	synccons();
+	int ms;
 
+	synccons();
 	/* accounting */
 	if(!m->online)
-		ndnr();
+		archreset();	/* exiting a second time; must not hang */
 	m->online = 0;
 	iprint("cpu%d: exiting\n", m->machno);
 	adec(&active.nonline);
@@ -397,7 +396,7 @@
 	/* wait (terminals wait forever) */
 	if(ispanic && !cpuserver)
 		ndnr();
-	for(int ms = 0; ms < 5000; ms += 1){
+	for(ms = 0; ms < 5000; ms += 1){
 		if(active.nonline == 0)
 			break;
 		delay(1);
--- /sys/src/nix/k10/rebootcode.s	Thu Jan  1 00:00:00 1970
+++ /sys/src/nix/k10/rebootcode.s	Sat Mar 22 02:02:07 2014
@@ -0,0 +1,111 @@
+#include "mem.h"
+#include "amd64l.h"
+
+MODE $64
+
+/*
+ * Turn off MMU, then memmove the new kernel to its correct location
+ * in physical memory.  Then jumps the to start of the kernel.
+ */
+
+TEXT	main(SB), 1, $-4
+	MOVL	RARG, DI		/* destination */
+	MOVL	p2+8(FP), SI		/* source */
+	MOVL	n+16(FP), BX		/* byte count */
+
+	/* load zero length idt */
+	MOVL	$_idtptr64p<>(SB), AX
+	MOVL	(AX), IDTR
+
+	/* load temporary gdt */
+	MOVL	$_gdtptr64p<>(SB), CX
+	MOVL	(CX), GDTR
+	/* move stack below destination */
+	MOVL	DI, SP
+
+	/* load CS with 32bit code segment */
+	PUSHQ	$SSEL(SiU32CS, 0)
+	PUSHQ	$_warp32<>(SB)
+	RETFQ
+//
+//TEXT	wave(SB), 1, $-4
+//	MOVL $0x3f8, DX
+//	MOVB $'X', AL
+//	OUTB
+//	MFENCE
+//	RET
+
+MODE $32
+
+TEXT	_warp32<>(SB), 1, $-4
+
+	/* load 32bit data segments */
+	MOVL	$SSEL(2, 0), AX
+	MOVW	AX, DS
+	MOVW	AX, ES
+	MOVW	AX, FS
+	MOVW	AX, GS
+	MOVW	AX, SS
+
+	/* turn off paging */
+	MOVL	CR0, AX
+	ANDL	$0x7fffffff, AX		/* ~(PG) */
+	MOVL	AX, CR0
+
+	MOVL	$0, AX
+	MOVL	AX, CR3
+
+	/* disable long mode */
+	MOVL	$0xc0000080, CX		/* Extended Feature Enable */
+	RDMSR
+	ANDL	$0xfffffeff, AX		/* Long Mode Disable */
+	WRMSR
+
+	/* diable pae */
+	MOVL	CR4, AX
+	ANDL	$0xffffff5f, AX		/* ~(PAE|PGE) */
+	MOVL	AX, CR4
+
+	MOVL	BX, CX			/* byte count */
+	MOVL	DI, AX			/* save entry point */
+
+/*
+ * the source and destination may overlap.
+ * determine whether to copy forward or backwards
+ */
+	CMPL	SI, DI
+	JGT	_forward
+	MOVL	SI, DX
+	ADDL	CX, DX
+	CMPL	DX, DI
+	JGT	_back
+
+_forward:
+	CLD
+	REP;	MOVSB
+
+_startkernel:
+	JMP*	AX
+
+_back:
+	ADDL	CX, DI
+	ADDL	CX, SI
+	SUBL	$1, DI
+	SUBL	$1, SI
+	STD
+	REP;	MOVSB
+	JMP	_startkernel
+
+TEXT _gdt<>(SB), 1, $-4
+	QUAD	$0
+	QUAD	$(SdL|SdG|SdP|Sd4G|SdDPL0|SdS|SdCODE|SdR)	/* 64-bit executable */
+	QUAD	$(SdG|SdD|Sd4G|SdP|SdDPL0|SdS|SdW)			/* 32-bit data */
+	QUAD	$(SdG|SdD|Sd4G|SdP|SdDPL0|SdS|SdCODE|SdW)	/* 32-bit exec */
+
+TEXT _gdtptr64p<>(SB), 1, $-4
+	WORD	$(4*8-1)
+	QUAD	$_gdt<>(SB)
+
+TEXT _idtptr64p<>(SB), 1, $-4
+	WORD	$0
+	QUAD	$0