//-------------------------------------------------------------------
//	xhcidemo.cpp
//
//	This example illustrates direct hardware programming of the
//	USB 3.0 Extensible Host Controller on our PCIe adapter-card
//	when our Delcom USB Visual Indicator is attached to a port:
//	its blue lamp remains illuminated until <ENTER>-key is hit.
//	
//	  to compile:  $ g++ xhcidemo.cpp -o xhcidemo
//	  to execute:  $ ./xhcidemo
//
//	NOTE: Our 'usb3.c' kernel-module needs to be pre-installed.
//
//	programmer: ALLAN CRUSE
//	date begun: 09 OCT 2010
//	completion: 13 OCT 2010
//	revised on: 03 DEC 2010 -- to improve event-ring's display 
//-------------------------------------------------------------------

#include <stdio.h>	// for printf(), perror() 
#include <fcntl.h>	// for open() 
#include <stdlib.h>	// for exit() 
#include <unistd.h>	// for lseek() 
#include <string.h>	// for memset()
#include <signal.h>	// for signal()
#include <sys/mman.h>	// for mmap()
#include <sys/ioctl.h>	// for ioctl()

#define XHCIBASE    0x80000	// suitable address for io-mapping
#define XHCISIZE    0x02000	// length of xHCI's register-space
#define SLOTS_EN	  2	// number of device-slots enabled


typedef struct	{
		unsigned long long	data;
		unsigned int		stat;
		unsigned int		ctrl;
		} TRB;	// Transfer Request Block

typedef struct	{ 
		unsigned long long  	addr;
		unsigned long long	elts; 
		} STE;	// Segment Table Entry


unsigned int	*lp = (unsigned int *)XHCIBASE;	// capabilities
unsigned int	*op, *rt, *db;	// operationals, runtimes, doorbells
unsigned int	n_slots;	// maximum number of device slots 
unsigned int	n_intrs;	// maximum number of interrupters
unsigned int	n_ports;	// maxcimum number of ports
char  devname[] = "/dev/xhci";	// name of driver device-node
unsigned long	kmem_phys = 0;	// physical address of memory arena
unsigned char	*kmem = (unsigned char *)( XHCIBASE + XHCISIZE );
unsigned int	*dc1 = (unsigned int *)( kmem + 0x0000 );
unsigned int	*dc2 = (unsigned int *)( kmem + 0x0800 );
unsigned long long	*dcbaa = (unsigned long long *)( kmem + 0x1000 );
unsigned long	dc1_phys, dc2_phys, dcbaa_phys, inctx_phys;
unsigned int	*ic  = (unsigned int *)( kmem + 0x2000 );
TRB		*cring = (TRB *)( kmem + 0x2800 );
TRB		*ering = (TRB *)( kmem + 0x3000 );
STE		*erstb = (STE *)( kmem + 0x3800 );
TRB		*tring = (TRB *)( kmem + 0x4000 );
unsigned long	cring_phys, ering_phys, erstb_phys;
unsigned long	tring_phys, dd_phys, delc_phys;
unsigned char	*dd = (unsigned char *)( kmem + 0x5000 );
unsigned long long	*delcom = (unsigned long long *)( kmem + 0x6000 );
 

void show_operational_registers( void );
void show_port_status_registers( void );
void show_runtime_registers( void );
void show_device_control_base_address_array( unsigned long long * );
void show_input_device_context( void );
void show_event_ring_entries( void );
void show_command_ring_entries( void );
void reset_xhci_controller( void );
void my_sigint_handler( int signo ) { exit(1); } 


int main( int argc, char **argv )
{
	// open the device-file for reading and writing
	int	xd = open( devname,O_RDWR );
	if ( xd < 0 ) { perror( devname ); exit(1); }

	// map the XHCI registers to user-space
	int	size = lseek( xd, 0, SEEK_END );
	int	prot = PROT_READ | PROT_WRITE;
	int	flag = MAP_FIXED | MAP_SHARED;
	void	*mm = (void *)XHCIBASE;
	if ( mmap( mm, size, prot, flag, xd, 0 ) == MAP_FAILED )
		{ perror( "mmap" ); exit(1); }

	// initialize our kernel memory arena
	memset( kmem, 0x00, size - XHCISIZE );

	// initialize our pointers to XHCI registers
	op = (unsigned int *)( XHCIBASE + (lp[ 0x00 >> 2 ] & 0xFF) );
	rt = (unsigned int *)( XHCIBASE + lp[ 0x18 >> 2 ] );
	db = (unsigned int *)( XHCIBASE + lp[ 0x14 >> 2 ] );

	// insure xhci controller gets reset upon exit
	atexit( reset_xhci_controller );
	signal( SIGINT, my_sigint_handler );

	// setup our kernel-memory data-structures
	if ( ioctl( xd, 0, &kmem_phys ) < 0 ) { perror( "ioctl" ); exit(1); }
	dc1_phys = kmem_phys + 0x0000;
	dc2_phys = kmem_phys + 0x0800;
	dcbaa[ 0 ] = 0;
	dcbaa[ 1 ] = dc1_phys;
	dcbaa[ 2 ] = dc2_phys;
	dcbaa_phys = kmem_phys + 0x1000;
	inctx_phys = kmem_phys + 0x2000;
	cring_phys = kmem_phys + 0x2800;
	ering_phys = kmem_phys + 0x3000;
	erstb_phys = kmem_phys + 0x3800;
	erstb[0].addr = ering_phys;
	erstb[0].elts = 32;
	tring_phys = kmem_phys + 0x4000;
	dd_phys = kmem_phys + 0x5000;
	delc_phys = kmem_phys +0x6000;

	// initialize our counters
	unsigned int	hcsparams1 = lp[ 0x04 >> 2 ];
	n_slots = (hcsparams1 >> 0) & 0x0FF;
	n_intrs = (hcsparams1 >> 8) & 0x7FF;
	n_ports = (hcsparams1 >>24) & 0x0FF; 

	// show initial register-state
	show_operational_registers();
	show_port_status_registers();
	show_runtime_registers();

	// show our data-structures
	show_device_control_base_address_array( &dcbaa[0] );
	show_input_device_context();
	

	// initialize the XHCI controller
	while ( op[ 0x04 >> 2 ] & (1 << 11) );	// spin until CNR==0
	op[ 0x38 >> 2 ] = SLOTS_EN;		// SlotsEnable = 2 
	op[ 0x30 >> 2 ] = dcbaa_phys;		// DCBAAP register
	op[ 0x18 >> 2 ] = cring_phys | 1;	// CRCR - low
	op[ 0x1C >> 2 ] = 0;			// CRCR - high
	rt[ 0x20 >> 2 ] = (1 << 1);		// IMAN IE-bit set
	rt[ 0x24 >> 2 ] = 0;			// IMOD no throttling 
	rt[ 0x28 >> 2 ] = 1;			// ERSTSZ = 1
	rt[ 0x38 >> 2 ] = ering_phys;		// ERSTDP start
	rt[ 0x30 >> 2 ] = erstb_phys;		// ERSTBA - low ebables!
	rt[ 0x34 >> 2 ] = 0;			// ERSTBA - high
	op[ 0x00 >> 2 ] = op[ 0x00 ] | (1<<0); 	// USBCMD Run/Stop = 1
	while ( op[ 0x04 >> 2 ] & (1 << 0 ) );	// USBSTS HCH==0?

	// setup 'Enable Slot' command
	cring[0].data = 0; 
	cring[0].stat = 0;
	cring[0].ctrl = (9<<10)|1;
	// ring the doorbell
	db[ 0 ] = 0;
	usleep( 1000 );

	// enable root ports
	for (int i = 0; i < n_ports; i++)
		{
		unsigned int	index = (0x400 + 16*i) >> 2;
		unsigned int	portsc = op[ index ];
		printf( "\n port #%X: ", i+1 );
		op[ index ] = portsc;		// clear WC-bits
		if ( (portsc & (1<<0)) == 0 ) 
			{ printf( " unconnected " ); continue; } 
		if ( (portsc & (1<<1)) != 0 )
			{ printf( " already enabled "); continue; }
		printf( " resetting... " );
		op[ index ] |= (1 << 4);	// set PR-bit
		// wait for PRC==1, PR==0, PED==1, PLS==0
		do	{
			portsc = op[ index ];
			}
		while ( ( portsc & 0x002001F2 ) != 0x00200002 );
		printf( "done " );
		}
	printf( "\n" );
	show_port_status_registers();
	show_runtime_registers();
	getchar();
	show_event_ring_entries();

	//----------------------------------------
	// Next is the 'Address_Device' operation
	//----------------------------------------

	// initialize input device context for 'Address Enable'
	ic[ 0 ] = 0;  ic[ 1 ] = 3;

	// Initialize the Input Slot Context
	unsigned int	*sc = (unsigned int *)&ic[8];
	unsigned int	portsc = op[ (0x400 + 2*16) >> 2 ];
	printf( "--- PORTSC%X=%08X --- \n\n", 3, portsc );
	unsigned int	speed = (portsc >> 10)&0xF;
	sc[0] = (1 << 27) | (speed << 20);	// context entries speed
	sc[1] = (3 << 16);

	// Initialize the Input Endpoint 0 Context
	unsigned int	*ec = (unsigned int *)&ic[16];
	ec[1] = (8 << 16) | (4 << 3) | (3 << 1);
	ec[2] = tring_phys | 1;		
	show_input_device_context();

	// Setup the 'Address_Device' command TRB
	unsigned int	dev = 1;
	cring[1].data = inctx_phys;
	cring[1].stat = 0;
	cring[1].ctrl = (dev<<24)|(11<<10)|1;
	// ring the doorbell
	db[0] = 0;
	usleep( 1000 );

	// setupthe Delcom Visual Indicator commands
	delcom[0] = 0x00040C65;		// turn on 'blue'
	delcom[1] = 0x04000C65;		// turn off 'blue'

	// setup the Transfer TRBs 
	tring[0].data = 0;
	tring[0].stat = 0;
	tring[0].ctrl = (8<<10)|1;		// 'No Op' command

	// 'Get_Device_Descriptor' control-transfer
	tring[1].data = 0x0012000001000680;
	tring[1].stat = 8;
	tring[1].ctrl = (3<<16)|(2<<10)|(1<<6)|1;	// SETUP

	tring[2].data = dd_phys;
	tring[2].stat = 18;
	tring[2].ctrl = (1<<16)|(3<<10)|(1<<2)|1;	// DATA

	tring[3].data = 0;
	tring[3].stat = 0;
	tring[3].ctrl = (0<<16)|(4<<10)|(1<<5)|1;	// HANDSHAKE

	// 'Set_Device_Configuration' control-transfer
	tring[4].data = 0x0000000000010900;
	tring[4].stat = 8;
	tring[4].ctrl = (0<<16)|(2<<10)|(1<<6)|1;	// SETUP	

	tring[5].data = 0;
	tring[5].stat = 0;
	tring[5].ctrl = (1<<16)|(4<<10)|(1<<5)|1;	// HANDSHAKE

	// 'Set_Interface_Configuration' control-transfer
	tring[6].data = 0x0008000000000921;
	tring[6].stat = 8;
	tring[6].ctrl = (2<<16)|(2<<10)|(1<<6)|1;	// SETUP

	tring[7].data = delc_phys + 0;
	tring[7].stat = 8;
	tring[7].ctrl = (0<<16)|(3<<10)|(1<<2)|1;	// DATA

	tring[8].data = 0; 
	tring[8].stat = 0;
	tring[8].ctrl = (1<<16)|(4<<10)|(1<<5)|1;	// HANDSHAKE

	// reserve space for a 'No Op' command
	tring[9].data = 0;
	tring[9].stat = 0;
	tring[9].ctrl = 0;

	// 'Set_Interface_Configuration' control-transfer
	tring[10].data = 0x0008000000000921;
	tring[10].stat = 8;
	tring[10].ctrl = (2<<16)|(2<<10)|(1<<6)|1;	// SETUP

	tring[11].data = delc_phys + 8;
	tring[11].stat = 8;
	tring[11].ctrl = (0<<16)|(3<<10)|(1<<2)|1;	// DATA

	tring[12].data = 0; 
	tring[12].stat = 0;
	tring[12].ctrl = (1<<16)|(4<<10)|(1<<5)|1;	// HANDSHAKE

	// ring the doorbell with target = 1
	db[1] = 1;				// ring doorbell
	usleep( 5000 );
	getchar();

	// turn off lamp 
	tring[9].ctrl = (8<<10)|1;		// 'No Op'
	db[1] = 1;				// ring doorbell
	usleep( 5000 );
	show_event_ring_entries();
	show_command_ring_entries();
}


void reset_xhci_controller( void )
{
	// stop the XHCI controller and reset it
	op[ 0x00 >> 2 ] = 0;			// USBCMD Run/Stop=0
	while ( (op[ 0x04 >> 2 ] & 1) == 0 );	// USBSTS HCH==1?
	op[ 0x00 >> 2 ] |= (1 << 1);		// USBCMD HCRST=1
	while ( (op[ 0x00 >> 2 ] & 2) != 0 );	// USBCMD HCRST==0?
}


void show_operational_registers( void )
{
	printf( "\n");
	printf( " USBCMD=%08X ", op[ 0x00 >> 2 ] );
	printf( " USBSTS=%08X ", op[ 0x04 >> 2 ] );
	printf( " PAGESIZE=%08X ", op[ 0x08 >> 2 ] );
	printf( " DNCTLOFF=%08X ", op[ 0x14 >> 2 ] );
	printf( "\n");
	printf( " CRCR=%08X%08X ", op[ 0x1C >> 2 ], op[ 0x18 >> 2 ] );
	printf( " DCBAAP=%08X%08X ", op[ 0x34 >> 2 ], op[ 0x30 >> 2 ] );
	printf( " CONFIG=0x%02X ", op[ 0x38 >> 2 ] );
	printf( "\n" );
}

void show_port_status_registers()
{
	int	i;

	for (i = 0; i < n_ports; i++)
		{
		unsigned int	index = 0x400 + 16*i;
		unsigned int	portsc = op[ index >> 2 ];
		printf( "\n" );
		printf( " PORTSC%X=%08X ", i+1, portsc );
		printf( " speed=%X ", (portsc >> 10)&0xF );
		printf( " powered=%X ", (portsc >> 9)&1 );
		printf( " state=%X ", (portsc >> 5)&0xF );
		printf( " enabled=%X ", (portsc >> 1)&1 );
		printf( " connected=%X ", (portsc >> 0)&1 );
		}
	printf( "\n" );
}
	
void show_runtime_registers( void )
{
	int	i;

	printf( "\n");
	printf( " MFINDEX=%04X ", rt[ 0x00 >> 2 ] );
	printf( "  IMAN      IMOD     ERSTSZ      " );
	printf( "  ERSTBA             ERDP  " );
	for (i = 0; i < n_intrs; i++)
		{
		unsigned 	index = 0x20*(i+1);
		unsigned int	iman, imod, erstsz;
		unsigned long long	erstba, erdp;
		iman = rt[ (index + 0x00) >> 2 ];
		imod = rt[ (index + 0x04) >> 2 ];
		erstsz = rt[ (index + 0x08) >> 2 ];
		erstba = rt[ (index + 0x10) >> 2 ];
		erdp = rt[ (index + 0x18) >> 2 ];
		printf( "\n    intr #%X  ", i );
		printf( " %08X ", iman );
		printf( " %08X ", imod );
		printf( " %08X ", erstsz );
		printf( " %016X ", erstba );
		printf( " %016X ", erdp );
		}	
	printf( "\n");

	printf( "\n");
}
	
void show_device_control_base_address_array( unsigned long long *dcbaap )
{
	int	dcbaa_entries = 1 + ((1 + SLOTS_EN) | 7);
	printf( " Device Context Base Address Array: " );
	for (int i = 0; i < dcbaa_entries; i++)
		{
		if ( (i % 4) == 0 ) printf( "\n  DC%02d: ", i );
		printf( " %016llX", dcbaap[ i ] );
		}
	printf( "\n" );
	printf( "\n" );
}

void show_input_device_context( void )
{
	printf( " Input Device Context: " );
	printf( "\n" );
	printf( "    Input Control Context: " );
	printf( " DROP=%08X  ADD=%08X \n", ic[0], ic[1] );	
	printf( "       Input Slot Context: " );
	printf( " %08X %08X %08X %08X \n", ic[8], ic[9], ic[10], ic[11] );
	printf( " Input Endpoint 0 Context: " );
	printf( " %08X %08X %08X%08X ", ic[16], ic[17], ic[19], ic[18] );
	printf( "BIDIRECTIONAL\n" );
	for (int i = 2; i < 32; i++)
		{
		int	endpt = i/2;
		int	index = (i+1) * 8;
		int	state = ic[ index ] & 7;
		if ( state == 0 ) continue;
		printf( " Input Endpoint %X Context: ", endpt );
		printf( " %08X %08X", ic[ index + 0 ], ic[ index + 1 ] );
		printf( " %08X%08X ", ic[ index + 3 ], ic[ index + 2 ] );
		switch ( i%2 )
			{
			case 0:	printf( "OUT \n" ); break;
			case 1: printf( "IN  \n" ); break;
			} 
		}
	printf( "\n" );
}


void show_event_ring_entries( void )
{
	printf( " Event Ring for Primary Interrupter: " );
	for (int i = 0; i < 32; i++)
		{
		unsigned long	trb_phys = ering_phys + (i * 16); 
		unsigned int	*ep = (unsigned int *)&ering[ i ];
		unsigned int	trbtype = (ep[3] >> 10)&0x3F;
		printf( "\n <0x%08lX>: ", trb_phys );
		for (int j = 0; j < 4; j++) printf( "%08X ", ep[j] );
		switch ( trbtype )
			{
			case  0: break;
			case 32: printf( "Transfer " ); break;
			case 33: printf( "Command Completion " ); break;
			case 34: printf( "Port Status Change " ); break;
			default: printf( "--- ? --- " ); break;
			}
		if ( trbtype == 0 ) break;
		}
	printf( "\n\n" );
}

void show_command_ring_entries( void )
{
	printf( " Command Ring entrires: " );
	for (int i = 0; i < 32; i++)
		{
		unsigned long	trb_phys = cring_phys + (i * 16); 
		unsigned int	*ep = (unsigned int *)&cring[ i ];
		unsigned int	trbtype = (ep[3] >> 10)&0x3F;
		printf( "\n <0x%08lX>: ", trb_phys );
		for (int j = 0; j < 4; j++) printf( "%08X ", ep[j] );
		switch ( trbtype )
			{
			case  0: break;
			case  9: printf( "Enable Slot " ); break;
			case 11: printf( "Address Device " ); break;
			default: printf( "--- ? --- " ); break;
			}
		if ( trbtype == 0 ) break;
		}
	printf( "\n\n" );
}