/*
 * 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 "regs4740.h"
#include "ubb-vga.h"


#define	REG_BASE_PTR	base

static volatile void *base;

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	DMA		0


static uint32_t old_icmr;
static uint32_t old_clkgr;


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(TIMER) = 1;	/* count at PCLK/1 */
	TDFR(TIMER) = 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)
{
	base = map(SOC_BASE, REG_WINDOW);

	/*
	 * 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(TIMER) & 0xffff) < cycles);
}


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


static const struct mode {
	const char *name;
	int xres, yres;
	int line_words;		/* xres/8 */
	int clkdiv;		/* pixel clock = 336 MHz/(clkdiv+1) */
	int line_cycles;	/* 31.77 us for official VGA */
	int hsync_end;		/* 0.79+3.77 us for official VGA */
} mode_db[] = {
    { "640x480",   640, 480,  640/8, 11, US(29.7), US(0.79+3.77-0.3) },
    { "800x600",   800, 600,  800/8,  8, US(28.7), US(2.0+3.3+0.3) },
    /* the next one may work after adjusting the timing in "frame" */
    { "800x600",   800, 600,  800/8,  8, US(28.2), US(2.0+3.3+0.3-0.3) },
    /* the 1024x768 below is not great but has good parameter tolerance */
    { "1024x768", 1024, 768, 1024/8,  8, US(36.0), US(2.0+3.3) },
    /* illustrate underruns */
    { "1024x768ur", 1024, 768, 1024/8,  7, US(33.5), US(0.4+2.1+0.5) },
    { NULL }
}, *mode = mode_db;

	/*
	 * Adjustment value tests with the XEN-1510 (640x480):
	 *
	 * 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
	 */


static 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 = mode->clkdiv;	/* 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 setup_noirq(void)
{
	DMAC = 1;
	DCS(DMA) = (1 << 3) | (1 << 2);
	DCS(DMA) = 0;

	DCM(DMA) =
	    (1 << 23) |         /* source address increment */
	    (4 << 8);           /* transfer size is 32 bytes */
	DRT(DMA) = 26;          /* MSC transmit-fifo-empty transfer request */
}


static void cleanup_noirq(void)
{
	DMAC = 0;
	DCS(DMA) = (1 << 3) | (1 << 2);
	DCS(DMA) = 0;
}


static void line(unsigned long line)
{
	/* Back porch */

	MSC_STRPCL = 1 << 3;	/* reset the MSC */

//	DCS(DMA) = (1 << 31) | (1 << 3) | (1 << 2);
	DCS(DMA) = 1 << 31;
	DSA(DMA) = line;
	DTA(DMA) = REG_PADDR(MSC_TXFIFO);	/* MUST set this each time */
	DTC(DMA) = mode->xres >> 6;

	until(US(0.79));

	/* HSYNC */

	PDDATC = HSYNC;
	MSC_STRPCL = 2;		/* start MMC clock output */
	MSC_RESTO = 0xffff;

	MSC_CMDAT =
	    (2 << 9) |		/* 4 bit bus */
	    (1 << 8) |		/* DMA */
	    (1 << 4) |		/* write */
	    (1 << 3) |		/* with data transfer */
	    1;			/* R1 response */

	MSC_STRPCL = 4;		/* START_OP */

	DCS(DMA) =
	    (1 << 31) |         /* no descriptor */
	    1;           

	until(mode->hsync_end);

//	MSC_TXFIFO = 0xffffffff;

	/* Front porch */

	PDFUNS = CMD;
	PDDATS = HSYNC;
	PDFUNC = CMD;

	until(mode->line_cycles);
	if (MSC_STAT & 3)
		bad++;
}


static void hdelay(int cycles)
{
	while (cycles--) {
		TCNT(TIMER) = 0;
		PDDATC = HSYNC;
		until(US(3.77));
		PDDATS = HSYNC;
		until(mode->line_cycles);
	}
}


static void frame(const unsigned long *f)
{
	const unsigned long *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(TIMER) = 0;
	PDDATC = HSYNC;
	until(US(3.77));
	PDDATS = HSYNC;
	until(mode->line_cycles-US(0.79));

	/*
	 * Note: resetting the timer just before calling "line" isn't enough.
	 * We have t reset it before the loop and right after returning from
	 * "line".
	 */
	TCNT(TIMER) = 0;
	for (p = f; p != f+mode->yres; p++) {
		line(*p);
		TCNT(TIMER) = 0;
	}

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


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


static void session(void (*gen)(void **fb, int xres, int yres), int frames)
{
	void **f_virt;
	const unsigned long *f_phys;
	int i;

	ccube_init();
	f_virt = calloc_phys_vec(mode->yres, mode->xres/2);
	gen(f_virt, mode->xres, mode->yres);
	f_phys = xlat_virt(f_virt, mode->yres);

	disable_interrupts();

	setup_noirq();

	for (i = 0; i != frames; i++) {
		frame(f_phys);
		if (DTC(DMA)) {
			fprintf(stderr,
			    "DMA locked up. Need hardware reset.\n");
			break;
		}
	}

	cleanup_noirq();

	enable_interrupts();
}


static void usage(const char *name)
{
	fprintf(stderr,
"usage: %s [-t] [-r resolution] frames [file]\n\n"
"  frames  number of frames to display\n"
"  file    PPM file\n\n"
"  -m mode  select the display mode, default \"%s\"\n"
"  -t       generate a test image\n"
    , name, mode_db[0].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, "m:t")) != EOF)
		switch (c) {
		case 'm':
			for (mode = mode_db; mode->name; mode++)
				if (!strcmp(mode->name, optarg))
					break;
			if (!mode->name) {
				fprintf(stderr, "no mode \"%s\"\n", optarg);
				exit(1);
			}
			break;
		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;
}