/*
 * ubb-vga.c - Output video on UBB with more or less VGA timing
 *
 * Written 2011 by Werner Almesberger
 * Copyright 2011 Werner Almesberger
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 */

/*
 * WARNING: this program does very nasty things to the Ben and it doesn't
 * like company. In particular, it resents:
 *
 * - the MMC driver - disable it with
 *   echo jz4740-mmc.0 >/sys/bus/platform/drivers/jz4740-mmc/unbind
 * - the AT86RF230/1 kernel driver - use a kernel that doesn't have it
 * - anything that accesses the screen - kill GUI, X server, etc.
 * - the screen blanker - either disable it or make sure the screen stays
 *   dark, e.g., with
 *   echo 1 >/sys/devices/platform/jz4740-fb/graphics/fb0/blank
 * - probably a fair number of other daemons and things as well - best to
 *   kill them all.
 */


#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <string.h>
#include <fcntl.h>
#include <sys/mman.h>

#include "ubb-vga.h"


static int bad;

/* ----- I/O pin assignment ------------------------------------------------ */


#define	DAT0		(1 << 10)
#define	DAT1		(1 << 11)
#define	DAT2		(1 << 12)
#define	DAT3		(1 << 13)
#define	CMD		(1 << 8)
#define	CLK		(1 << 9)

#define	R		DAT3
#define	G		DAT0
#define	B		DAT1
#define	Y		DAT2
#define	HSYNC		CMD
#define	VSYNC		CLK


/* ----- Ben hardware ------------------------------------------------------ */


#define	TIMER		7


#define	PAGE_SIZE	4096
#define	SOC_BASE	0x10000000


static volatile uint32_t *icmr, *icmsr, *icmcr;
static uint32_t old_icmr;

static volatile uint32_t *clkgr, *msccdr;
static uint32_t old_clkgr;

static volatile uint32_t *pddats, *pddatc;
static volatile uint32_t *pddirs, *pddirc;
static volatile uint32_t *pdfuns, *pdfunc;

static volatile uint32_t *tssr, *tscr;
static volatile uint32_t *tesr, *tecr;
static volatile uint32_t *tcsr, *tdfr, *tcnt;

static volatile uint32_t *msc_strpcl, *msc_stat, *msc_clkrt;
static volatile uint32_t *msc_cmdat, *msc_resto, *msc_blklen, *msc_nob;
static volatile uint32_t *msc_cmd, *msc_arg, *msc_txfifo;


static void disable_interrupts(void)
{
	/*
	 * @@@ Race condition alert ! If we get interrupted/preempted between
	 * reading ICMR and masking all interrupts, and the code that runs
	 * between these two operations changes ICMR, then we may set an
	 * incorrect mask when restoring interrupts, which may hang the system.
	 */

	old_icmr = *icmr;
	*icmsr = 0xffffffff;
}


static void enable_interrupts(void)
{
	*icmcr = ~old_icmr;
}


/*
 * @@@ Disabling the LCD clock will hang operations that depend on the LCD
 * subsystem to advance. This includes the screen saver.
 */

static void disable_lcd(void)
{
	old_clkgr = *clkgr;
	*clkgr = old_clkgr | 1 << 10;
}


static void enable_lcd(void)
{
	*clkgr = old_clkgr;
}


static void get_timer(void)
{
	*tscr = 1 << TIMER;	/* enable clock */
	*tcsr = 1;		/* count at PCLK/1 */
	*tdfr = 0xffff;		/* count to 0xffff */
	*tesr = 1 << TIMER;
}


static void release_timer(void)
{
	*tecr = 1 << TIMER;
	*tssr = 1 << TIMER;
}


void *map(off_t addr, size_t size)
{
	int fd;
	void *mem;

	fd = open("/dev/mem", O_RDWR | O_SYNC);
	if (fd < 0) {
		perror("/dev/mem");
		exit(1);
	}
	mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, addr);
	if (mem == MAP_FAILED) {
		perror("mmap");
		exit(1);
	}

	return mem;
}


static void ben_setup(void)
{
	volatile void *base;

	base = map(SOC_BASE, PAGE_SIZE*3*16);

	clkgr		= base+0x20;
	msccdr		= base+0x68;

	icmr		= base+0x1004;
	icmsr		= base+0x1008;
	icmcr		= base+0x100c;

	pddats		= base+0x10314;
	pddatc		= base+0x10318;
	pdfuns		= base+0x10344;
	pdfunc		= base+0x10348;
	pddirs		= base+0x10364;
	pddirc		= base+0x10368;

	tssr		= base+0x202c;
	tscr		= base+0x203c;
	tesr		= base+0x2014;
	tecr		= base+0x2018;
	tcsr		= base+0x204c+0x10*TIMER;
	tdfr		= base+0x2040+0x10*TIMER;
	tcnt		= base+0x2048+0x10*TIMER;

	msc_strpcl	= base+0x21000;
	msc_stat	= base+0x21004;
	msc_clkrt	= base+0x21008;
	msc_cmdat	= base+0x2100c;
	msc_resto	= base+0x21010;
	msc_blklen	= base+0x21018;
	msc_nob		= base+0x2101c;
	msc_cmd		= base+0x2102c;
	msc_arg		= base+0x21030;
	msc_txfifo	= base+0x2103c;

	/*
	 * Ironically, switching the LCD clock on and off many times only
	 * increases the risk of a hang. Therefore, we leave stop it during
	 * all the measurements and only enable it again at the end.
	 */
	disable_lcd();
	get_timer();
}


static void cleanup(void)
{
	release_timer();
	enable_lcd();
}


/* ----- Delay logic ------------------------------------------------------- */


#define	US(us)	((uint16_t) ((us)*112))


static void until(uint16_t cycles)
{
	while ((*tcnt & 0xffff) < cycles);
}


/* ----- Frame buffer output ----------------------------------------------- */


static int line_words = 640/8;
//static int line_cycles = US(36); /* nominally 31.77 us, but we're too slow  */
//static int line_cycles = US(32); /* nominally 31.77 us, but we're too slow  */
static int line_cycles = US(29.6); /* nominally 31.77 us, but we're too fast */
/*
 * Note: 29.5 is already too short. Tricky timing.
 */

void setup(void)
{
	mlockall(MCL_CURRENT | MCL_FUTURE);
	ben_setup();

	*pdfuns = R | G | B | Y;
	*pdfunc = VSYNC | HSYNC;
	*pddirs = VSYNC | HSYNC | R | G | B | Y;
	*pddats = VSYNC | HSYNC;
	*pddatc = R | G | B | Y;

//	*msccdr = 20;		/* set the MSC clock to 336 MHz / 21 = 16 MHz */
	*msccdr = 11;		/* set the MSC clock to 336 MHz / 12 = 28 MHz */
	*clkgr &= ~(1 << 7);	/* enable MSC clock */
	*msc_clkrt = 0;		/* bus clock = MSC clock / 1 */
}


static void line(const uint32_t *line, volatile const uint32_t *prefetch)
{
	const uint32_t *p = line;

	/* Back porch */

	*tcnt = 0;

	*msc_strpcl = 1 << 3;	/* reset the MSC */
//	while (*msc_stat & (1 << 15));

	*msc_txfifo = *p++;

	until(US(0.79));

	/* HSYNC */

	*pddatc = HSYNC;
	*msc_strpcl = 2;	/* start MMC clock output */

	*msc_cmdat =
	    (1 << 10) |		/* 4 bit bus */
	    (1 << 4) |		/* write */
	    (1 << 3) |		/* with data transfer */
	    1;			/* R1 response */

	*msc_strpcl = 4;	/* START_OP */

	until(US(0.79+3.77-0.3));
	/*
	 * Adjustment value tests with the XEN-1510:
	 *
	 * Adjustment	Tries	Good	Jam	FIFO jitter
	 *						Quick load
	 * -0.0		10	3	7	0	n
	 * -0.1		10	5	5	0	n
	 * -0.2		10	6	4	0	n
	 * -0.3		10	7	3	0	n
	 *		10	5	5	0	y
	 * -0.4		10	1	0	9	n
	 * 		10	5	0	5	n	repeat
	 *		10	5	0	5	y
	 * -0.5		10	3	0	7	n
	 * 		10	7	0	3	y
	 * -1.0		 5	0	5	0
	 *
	 * Good = image is stable
	 * Jam = does not detect the signal properly, loss of HSYNC, artefacts,
	 *       or no image at all
	 * FIFO jitter = some lines get shifted by a "digital" amount
	 */

	/* Front porch */

	*pdfuns = CMD;
	*pddats = HSYNC;
	*pdfunc = CMD;

	/*
	 * We don't wait for the end of the front porch because the beginning
	 * of pixel data is determined by the MSC. Instead, we make good use
	 * of the delay to shovel bits into the MSC's FIFO.
	 */

#if 0	/* quick load */
	*msc_txfifo = *p++;
	*msc_txfifo = *p++;
#endif
	while (p != line+line_words) {
		uint8_t st;
		do {
			st = *msc_stat;
			if (st & 3) {
				bad++;
				goto fail;
			}
		}
		while (st & (1 << 7));
		*msc_txfifo = *p++;
	}

fail:
	(void) *prefetch;
	until(line_cycles);
}


static void hdelay(int cycles)
{
	while (cycles--) {
		*tcnt = 0;
		*pddatc = HSYNC;
		until(US(3.77));
		*pddats = HSYNC;
		until(line_cycles);
	}
}


static void frame(const uint32_t *f)
{
	const uint32_t *p;

	/* VSYNC */
	*pddatc = VSYNC;
	hdelay(2);
	*pddats = VSYNC;

	/* Front porch */

	hdelay(31);

	/*
	 * The horizontal back porch of the previous line is handled inside
	 * "line", so we have to wait for less than a full line here.
	 */
	*tcnt = 0;
	*pddatc = HSYNC;
	until(US(3.77));
	*pddats = HSYNC;
	until(line_cycles-US(0.79));

	for (p = f; p != f+480*line_words; p += line_words)
		line(p, p+line_words);

	/* Back porch */
	hdelay(14);
}


/* ----- Command-line parsing and main loop -------------------------------- */


static void session(void (*gen)(void *fb, int xres, int yres), int frames)
{
	uint32_t f[2*240*(line_words+1)];
	int i;

	memset(f, 0, sizeof(f));
	ccube_init();
	gen(f, 640, 480);

	disable_interrupts();

	for (i = 0; i != frames; i++)
		frame(f);

	enable_interrupts();
}


static void usage(const char *name)
{
	fprintf(stderr,
"usage: %s [-t] frames [file]\n\n"
"  frames  number of frames to display\n"
"  file    PPM file\n\n"
"  -t            generate a test image\n"
    , name);
	exit(1);
}


int main(int argc, char *const *argv)
{
	void (*gen)(void *fb, int xres, int yres) = grabfb;
	int frames;
	int c;

	while ((c = getopt(argc, argv, "t")) != EOF)
		switch (c) {
		case 't':
			gen = tstimg;
			break;
		default:
			usage(*argv);
		}

	switch (argc-optind) {
	case 2:
		img_name = argv[optind+1];
		gen = ppmimg;
		/* fall through */
	case 1:
		frames = atoi(argv[optind]);
		break;
	default:
		usage(*argv);
	}

	setup();
	session(gen, frames);
	cleanup();

	if (bad)
		printf("%d timeout%s\n", bad, bad == 1 ? "" : "s");
	return 0;
}