xref: /openbmc/linux/drivers/scsi/mesh.c (revision 1da177e4)

/*
 * SCSI low-level driver for the MESH (Macintosh Enhanced SCSI Hardware)
 * bus adaptor found on Power Macintosh computers.
 * We assume the MESH is connected to a DBDMA (descriptor-based DMA)
 * controller.
 *
 * Paul Mackerras, August 1996.
 * Copyright (C) 1996 Paul Mackerras.
 *
 * Apr. 21 2002  - BenH		Rework bus reset code for new error handler
 *                              Add delay after initial bus reset
 *                              Add module parameters
 *
 * Sep. 27 2003  - BenH		Move to new driver model, fix some write posting
 *				issues
 * To do:
 * - handle aborts correctly
 * - retry arbitration if lost (unless higher levels do this for us)
 * - power down the chip when no device is detected
 */
#include <linux/config.h>
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/delay.h>
#include <linux/types.h>
#include <linux/string.h>
#include <linux/slab.h>
#include <linux/blkdev.h>
#include <linux/proc_fs.h>
#include <linux/stat.h>
#include <linux/interrupt.h>
#include <linux/reboot.h>
#include <linux/spinlock.h>
#include <asm/dbdma.h>
#include <asm/io.h>
#include <asm/pgtable.h>
#include <asm/prom.h>
#include <asm/system.h>
#include <asm/irq.h>
#include <asm/hydra.h>
#include <asm/processor.h>
#include <asm/machdep.h>
#include <asm/pmac_feature.h>
#include <asm/pci-bridge.h>
#include <asm/macio.h>

#include <scsi/scsi.h>
#include <scsi/scsi_cmnd.h>
#include <scsi/scsi_device.h>
#include <scsi/scsi_host.h>

#include "mesh.h"

#if 1
#undef KERN_DEBUG
#define KERN_DEBUG KERN_WARNING
#endif

MODULE_AUTHOR("Paul Mackerras (paulus@samba.org)");
MODULE_DESCRIPTION("PowerMac MESH SCSI driver");
MODULE_LICENSE("GPL");

static int sync_rate = CONFIG_SCSI_MESH_SYNC_RATE;
static int sync_targets = 0xff;
static int resel_targets = 0xff;
static int debug_targets = 0;	/* print debug for these targets */
static int init_reset_delay = CONFIG_SCSI_MESH_RESET_DELAY_MS;

module_param(sync_rate, int, 0);
MODULE_PARM_DESC(sync_rate, "Synchronous rate (0..10, 0=async)");
module_param(sync_targets, int, 0);
MODULE_PARM_DESC(sync_targets, "Bitmask of targets allowed to set synchronous");
module_param(resel_targets, int, 0);
MODULE_PARM_DESC(resel_targets, "Bitmask of targets allowed to set disconnect");
module_param(debug_targets, int, 0644);
MODULE_PARM_DESC(debug_targets, "Bitmask of debugged targets");
module_param(init_reset_delay, int, 0);
MODULE_PARM_DESC(init_reset_delay, "Initial bus reset delay (0=no reset)");

static int mesh_sync_period = 100;
static int mesh_sync_offset = 0;
static unsigned char use_active_neg = 0;  /* bit mask for SEQ_ACTIVE_NEG if used */

#define ALLOW_SYNC(tgt)		((sync_targets >> (tgt)) & 1)
#define ALLOW_RESEL(tgt)	((resel_targets >> (tgt)) & 1)
#define ALLOW_DEBUG(tgt)	((debug_targets >> (tgt)) & 1)
#define DEBUG_TARGET(cmd)	((cmd) && ALLOW_DEBUG((cmd)->device->id))

#undef MESH_DBG
#define N_DBG_LOG	50
#define N_DBG_SLOG	20
#define NUM_DBG_EVENTS	13
#undef	DBG_USE_TB		/* bombs on 601 */

struct dbglog {
	char	*fmt;
	u32	tb;
	u8	phase;
	u8	bs0;
	u8	bs1;
	u8	tgt;
	int	d;
};

enum mesh_phase {
	idle,
	arbitrating,
	selecting,
	commanding,
	dataing,
	statusing,
	busfreeing,
	disconnecting,
	reselecting,
	sleeping
};

enum msg_phase {
	msg_none,
	msg_out,
	msg_out_xxx,
	msg_out_last,
	msg_in,
	msg_in_bad,
};

enum sdtr_phase {
	do_sdtr,
	sdtr_sent,
	sdtr_done
};

struct mesh_target {
	enum sdtr_phase sdtr_state;
	int	sync_params;
	int	data_goes_out;		/* guess as to data direction */
	struct scsi_cmnd *current_req;
	u32	saved_ptr;
#ifdef MESH_DBG
	int	log_ix;
	int	n_log;
	struct dbglog log[N_DBG_LOG];
#endif
};

struct mesh_state {
	volatile struct	mesh_regs __iomem *mesh;
	int	meshintr;
	volatile struct	dbdma_regs __iomem *dma;
	int	dmaintr;
	struct	Scsi_Host *host;
	struct	mesh_state *next;
	struct scsi_cmnd *request_q;
	struct scsi_cmnd *request_qtail;
	enum mesh_phase phase;		/* what we're currently trying to do */
	enum msg_phase msgphase;
	int	conn_tgt;		/* target we're connected to */
	struct scsi_cmnd *current_req;		/* req we're currently working on */
	int	data_ptr;
	int	dma_started;
	int	dma_count;
	int	stat;
	int	aborting;
	int	expect_reply;
	int	n_msgin;
	u8	msgin[16];
	int	n_msgout;
	int	last_n_msgout;
	u8	msgout[16];
	struct dbdma_cmd *dma_cmds;	/* space for dbdma commands, aligned */
	dma_addr_t dma_cmd_bus;
	void	*dma_cmd_space;
	int	dma_cmd_size;
	int	clk_freq;
	struct mesh_target tgts[8];
	struct macio_dev *mdev;
	struct pci_dev* pdev;
#ifdef MESH_DBG
	int	log_ix;
	int	n_log;
	struct dbglog log[N_DBG_SLOG];
#endif
};

/*
 * Driver is too messy, we need a few prototypes...
 */
static void mesh_done(struct mesh_state *ms, int start_next);
static void mesh_interrupt(int irq, void *dev_id, struct pt_regs *ptregs);
static void cmd_complete(struct mesh_state *ms);
static void set_dma_cmds(struct mesh_state *ms, struct scsi_cmnd *cmd);
static void halt_dma(struct mesh_state *ms);
static void phase_mismatch(struct mesh_state *ms);


/*
 * Some debugging & logging routines
 */

#ifdef MESH_DBG

static inline u32 readtb(void)
{
	u32 tb;

#ifdef DBG_USE_TB
	/* Beware: if you enable this, it will crash on 601s. */
	asm ("mftb %0" : "=r" (tb) : );
#else
	tb = 0;
#endif
	return tb;
}

static void dlog(struct mesh_state *ms, char *fmt, int a)
{
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];
	struct dbglog *tlp, *slp;

	tlp = &tp->log[tp->log_ix];
	slp = &ms->log[ms->log_ix];
	tlp->fmt = fmt;
	tlp->tb = readtb();
	tlp->phase = (ms->msgphase << 4) + ms->phase;
	tlp->bs0 = ms->mesh->bus_status0;
	tlp->bs1 = ms->mesh->bus_status1;
	tlp->tgt = ms->conn_tgt;
	tlp->d = a;
	*slp = *tlp;
	if (++tp->log_ix >= N_DBG_LOG)
		tp->log_ix = 0;
	if (tp->n_log < N_DBG_LOG)
		++tp->n_log;
	if (++ms->log_ix >= N_DBG_SLOG)
		ms->log_ix = 0;
	if (ms->n_log < N_DBG_SLOG)
		++ms->n_log;
}

static void dumplog(struct mesh_state *ms, int t)
{
	struct mesh_target *tp = &ms->tgts[t];
	struct dbglog *lp;
	int i;

	if (tp->n_log == 0)
		return;
	i = tp->log_ix - tp->n_log;
	if (i < 0)
		i += N_DBG_LOG;
	tp->n_log = 0;
	do {
		lp = &tp->log[i];
		printk(KERN_DEBUG "mesh log %d: bs=%.2x%.2x ph=%.2x ",
		       t, lp->bs1, lp->bs0, lp->phase);
#ifdef DBG_USE_TB
		printk("tb=%10u ", lp->tb);
#endif
		printk(lp->fmt, lp->d);
		printk("\n");
		if (++i >= N_DBG_LOG)
			i = 0;
	} while (i != tp->log_ix);
}

static void dumpslog(struct mesh_state *ms)
{
	struct dbglog *lp;
	int i;

	if (ms->n_log == 0)
		return;
	i = ms->log_ix - ms->n_log;
	if (i < 0)
		i += N_DBG_SLOG;
	ms->n_log = 0;
	do {
		lp = &ms->log[i];
		printk(KERN_DEBUG "mesh log: bs=%.2x%.2x ph=%.2x t%d ",
		       lp->bs1, lp->bs0, lp->phase, lp->tgt);
#ifdef DBG_USE_TB
		printk("tb=%10u ", lp->tb);
#endif
		printk(lp->fmt, lp->d);
		printk("\n");
		if (++i >= N_DBG_SLOG)
			i = 0;
	} while (i != ms->log_ix);
}

#else

static inline void dlog(struct mesh_state *ms, char *fmt, int a)
{}
static inline void dumplog(struct mesh_state *ms, int tgt)
{}
static inline void dumpslog(struct mesh_state *ms)
{}

#endif /* MESH_DBG */

#define MKWORD(a, b, c, d)	(((a) << 24) + ((b) << 16) + ((c) << 8) + (d))

static void
mesh_dump_regs(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	volatile struct dbdma_regs __iomem *md = ms->dma;
	int t;
	struct mesh_target *tp;

	printk(KERN_DEBUG "mesh: state at %p, regs at %p, dma at %p\n",
	       ms, mr, md);
	printk(KERN_DEBUG "    ct=%4x seq=%2x bs=%4x fc=%2x "
	       "exc=%2x err=%2x im=%2x int=%2x sp=%2x\n",
	       (mr->count_hi << 8) + mr->count_lo, mr->sequence,
	       (mr->bus_status1 << 8) + mr->bus_status0, mr->fifo_count,
	       mr->exception, mr->error, mr->intr_mask, mr->interrupt,
	       mr->sync_params);
	while(in_8(&mr->fifo_count))
		printk(KERN_DEBUG " fifo data=%.2x\n",in_8(&mr->fifo));
	printk(KERN_DEBUG "    dma stat=%x cmdptr=%x\n",
	       in_le32(&md->status), in_le32(&md->cmdptr));
	printk(KERN_DEBUG "    phase=%d msgphase=%d conn_tgt=%d data_ptr=%d\n",
	       ms->phase, ms->msgphase, ms->conn_tgt, ms->data_ptr);
	printk(KERN_DEBUG "    dma_st=%d dma_ct=%d n_msgout=%d\n",
	       ms->dma_started, ms->dma_count, ms->n_msgout);
	for (t = 0; t < 8; ++t) {
		tp = &ms->tgts[t];
		if (tp->current_req == NULL)
			continue;
		printk(KERN_DEBUG "    target %d: req=%p goes_out=%d saved_ptr=%d\n",
		       t, tp->current_req, tp->data_goes_out, tp->saved_ptr);
	}
}


/*
 * Flush write buffers on the bus path to the mesh
 */
static inline void mesh_flush_io(volatile struct mesh_regs __iomem *mr)
{
	(void)in_8(&mr->mesh_id);
}


/*
 * Complete a SCSI command
 */
static void mesh_completed(struct mesh_state *ms, struct scsi_cmnd *cmd)
{
	(*cmd->scsi_done)(cmd);
}


/* Called with  meshinterrupt disabled, initialize the chipset
 * and eventually do the initial bus reset. The lock must not be
 * held since we can schedule.
 */
static void mesh_init(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	volatile struct dbdma_regs __iomem *md = ms->dma;

	mesh_flush_io(mr);
	udelay(100);

	/* Reset controller */
	out_le32(&md->control, (RUN|PAUSE|FLUSH|WAKE) << 16);	/* stop dma */
	out_8(&mr->exception, 0xff);	/* clear all exception bits */
	out_8(&mr->error, 0xff);	/* clear all error bits */
	out_8(&mr->sequence, SEQ_RESETMESH);
	mesh_flush_io(mr);
	udelay(10);
	out_8(&mr->intr_mask, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	out_8(&mr->source_id, ms->host->this_id);
	out_8(&mr->sel_timeout, 25);	/* 250ms */
	out_8(&mr->sync_params, ASYNC_PARAMS);

	if (init_reset_delay) {
		printk(KERN_INFO "mesh: performing initial bus reset...\n");

		/* Reset bus */
		out_8(&mr->bus_status1, BS1_RST);	/* assert RST */
		mesh_flush_io(mr);
		udelay(30);			/* leave it on for >= 25us */
		out_8(&mr->bus_status1, 0);	/* negate RST */
		mesh_flush_io(mr);

		/* Wait for bus to come back */
		msleep(init_reset_delay);
	}

	/* Reconfigure controller */
	out_8(&mr->interrupt, 0xff);	/* clear all interrupt bits */
	out_8(&mr->sequence, SEQ_FLUSHFIFO);
	mesh_flush_io(mr);
	udelay(1);
	out_8(&mr->sync_params, ASYNC_PARAMS);
	out_8(&mr->sequence, SEQ_ENBRESEL);

	ms->phase = idle;
	ms->msgphase = msg_none;
}


static void mesh_start_cmd(struct mesh_state *ms, struct scsi_cmnd *cmd)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	int t, id;

	id = cmd->device->id;
	ms->current_req = cmd;
	ms->tgts[id].data_goes_out = cmd->sc_data_direction == DMA_TO_DEVICE;
	ms->tgts[id].current_req = cmd;

#if 1
	if (DEBUG_TARGET(cmd)) {
		int i;
		printk(KERN_DEBUG "mesh_start: %p ser=%lu tgt=%d cmd=",
		       cmd, cmd->serial_number, id);
		for (i = 0; i < cmd->cmd_len; ++i)
			printk(" %x", cmd->cmnd[i]);
		printk(" use_sg=%d buffer=%p bufflen=%u\n",
		       cmd->use_sg, cmd->request_buffer, cmd->request_bufflen);
	}
#endif
	if (ms->dma_started)
		panic("mesh: double DMA start !\n");

	ms->phase = arbitrating;
	ms->msgphase = msg_none;
	ms->data_ptr = 0;
	ms->dma_started = 0;
	ms->n_msgout = 0;
	ms->last_n_msgout = 0;
	ms->expect_reply = 0;
	ms->conn_tgt = id;
	ms->tgts[id].saved_ptr = 0;
	ms->stat = DID_OK;
	ms->aborting = 0;
#ifdef MESH_DBG
	ms->tgts[id].n_log = 0;
	dlog(ms, "start cmd=%x", (int) cmd);
#endif

	/* Off we go */
	dlog(ms, "about to arb, intr/exc/err/fc=%.8x",
	     MKWORD(mr->interrupt, mr->exception, mr->error, mr->fifo_count));
	out_8(&mr->interrupt, INT_CMDDONE);
	out_8(&mr->sequence, SEQ_ENBRESEL);
	mesh_flush_io(mr);
	udelay(1);

	if (in_8(&mr->bus_status1) & (BS1_BSY | BS1_SEL)) {
		/*
		 * Some other device has the bus or is arbitrating for it -
		 * probably a target which is about to reselect us.
		 */
		dlog(ms, "busy b4 arb, intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception,
			    mr->error, mr->fifo_count));
		for (t = 100; t > 0; --t) {
			if ((in_8(&mr->bus_status1) & (BS1_BSY | BS1_SEL)) == 0)
				break;
			if (in_8(&mr->interrupt) != 0) {
				dlog(ms, "intr b4 arb, intr/exc/err/fc=%.8x",
				     MKWORD(mr->interrupt, mr->exception,
					    mr->error, mr->fifo_count));
				mesh_interrupt(0, (void *)ms, NULL);
				if (ms->phase != arbitrating)
					return;
			}
			udelay(1);
		}
		if (in_8(&mr->bus_status1) & (BS1_BSY | BS1_SEL)) {
			/* XXX should try again in a little while */
			ms->stat = DID_BUS_BUSY;
			ms->phase = idle;
			mesh_done(ms, 0);
			return;
		}
	}

	/*
	 * Apparently the mesh has a bug where it will assert both its
	 * own bit and the target's bit on the bus during arbitration.
	 */
	out_8(&mr->dest_id, mr->source_id);

	/*
	 * There appears to be a race with reselection sometimes,
	 * where a target reselects us just as we issue the
	 * arbitrate command.  It seems that then the arbitrate
	 * command just hangs waiting for the bus to be free
	 * without giving us a reselection exception.
	 * The only way I have found to get it to respond correctly
	 * is this: disable reselection before issuing the arbitrate
	 * command, then after issuing it, if it looks like a target
	 * is trying to reselect us, reset the mesh and then enable
	 * reselection.
	 */
	out_8(&mr->sequence, SEQ_DISRESEL);
	if (in_8(&mr->interrupt) != 0) {
		dlog(ms, "intr after disresel, intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception,
			    mr->error, mr->fifo_count));
		mesh_interrupt(0, (void *)ms, NULL);
		if (ms->phase != arbitrating)
			return;
		dlog(ms, "after intr after disresel, intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception,
			    mr->error, mr->fifo_count));
	}

	out_8(&mr->sequence, SEQ_ARBITRATE);

	for (t = 230; t > 0; --t) {
		if (in_8(&mr->interrupt) != 0)
			break;
		udelay(1);
	}
	dlog(ms, "after arb, intr/exc/err/fc=%.8x",
	     MKWORD(mr->interrupt, mr->exception, mr->error, mr->fifo_count));
	if (in_8(&mr->interrupt) == 0 && (in_8(&mr->bus_status1) & BS1_SEL)
	    && (in_8(&mr->bus_status0) & BS0_IO)) {
		/* looks like a reselection - try resetting the mesh */
		dlog(ms, "resel? after arb, intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception, mr->error, mr->fifo_count));
		out_8(&mr->sequence, SEQ_RESETMESH);
		mesh_flush_io(mr);
		udelay(10);
		out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
		out_8(&mr->intr_mask, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
		out_8(&mr->sequence, SEQ_ENBRESEL);
		mesh_flush_io(mr);
		for (t = 10; t > 0 && in_8(&mr->interrupt) == 0; --t)
			udelay(1);
		dlog(ms, "tried reset after arb, intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception, mr->error, mr->fifo_count));
#ifndef MESH_MULTIPLE_HOSTS
		if (in_8(&mr->interrupt) == 0 && (in_8(&mr->bus_status1) & BS1_SEL)
		    && (in_8(&mr->bus_status0) & BS0_IO)) {
			printk(KERN_ERR "mesh: controller not responding"
			       " to reselection!\n");
			/*
			 * If this is a target reselecting us, and the
			 * mesh isn't responding, the higher levels of
			 * the scsi code will eventually time out and
			 * reset the bus.
			 */
		}
#endif
	}
}

/*
 * Start the next command for a MESH.
 * Should be called with interrupts disabled.
 */
static void mesh_start(struct mesh_state *ms)
{
	struct scsi_cmnd *cmd, *prev, *next;

	if (ms->phase != idle || ms->current_req != NULL) {
		printk(KERN_ERR "inappropriate mesh_start (phase=%d, ms=%p)",
		       ms->phase, ms);
		return;
	}

	while (ms->phase == idle) {
		prev = NULL;
		for (cmd = ms->request_q; ; cmd = (struct scsi_cmnd *) cmd->host_scribble) {
			if (cmd == NULL)
				return;
			if (ms->tgts[cmd->device->id].current_req == NULL)
				break;
			prev = cmd;
		}
		next = (struct scsi_cmnd *) cmd->host_scribble;
		if (prev == NULL)
			ms->request_q = next;
		else
			prev->host_scribble = (void *) next;
		if (next == NULL)
			ms->request_qtail = prev;

		mesh_start_cmd(ms, cmd);
	}
}

static void mesh_done(struct mesh_state *ms, int start_next)
{
	struct scsi_cmnd *cmd;
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];

	cmd = ms->current_req;
	ms->current_req = NULL;
	tp->current_req = NULL;
	if (cmd) {
		cmd->result = (ms->stat << 16) + cmd->SCp.Status;
		if (ms->stat == DID_OK)
			cmd->result += (cmd->SCp.Message << 8);
		if (DEBUG_TARGET(cmd)) {
			printk(KERN_DEBUG "mesh_done: result = %x, data_ptr=%d, buflen=%d\n",
			       cmd->result, ms->data_ptr, cmd->request_bufflen);
			if ((cmd->cmnd[0] == 0 || cmd->cmnd[0] == 0x12 || cmd->cmnd[0] == 3)
			    && cmd->request_buffer != 0) {
				unsigned char *b = cmd->request_buffer;
				printk(KERN_DEBUG "buffer = %x %x %x %x %x %x %x %x\n",
				       b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7]);
			}
		}
		cmd->SCp.this_residual -= ms->data_ptr;
		mesh_completed(ms, cmd);
	}
	if (start_next) {
		out_8(&ms->mesh->sequence, SEQ_ENBRESEL);
		mesh_flush_io(ms->mesh);
		udelay(1);
		ms->phase = idle;
		mesh_start(ms);
	}
}

static inline void add_sdtr_msg(struct mesh_state *ms)
{
	int i = ms->n_msgout;

	ms->msgout[i] = EXTENDED_MESSAGE;
	ms->msgout[i+1] = 3;
	ms->msgout[i+2] = EXTENDED_SDTR;
	ms->msgout[i+3] = mesh_sync_period/4;
	ms->msgout[i+4] = (ALLOW_SYNC(ms->conn_tgt)? mesh_sync_offset: 0);
	ms->n_msgout = i + 5;
}

static void set_sdtr(struct mesh_state *ms, int period, int offset)
{
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	int v, tr;

	tp->sdtr_state = sdtr_done;
	if (offset == 0) {
		/* asynchronous */
		if (SYNC_OFF(tp->sync_params))
			printk(KERN_INFO "mesh: target %d now asynchronous\n",
			       ms->conn_tgt);
		tp->sync_params = ASYNC_PARAMS;
		out_8(&mr->sync_params, ASYNC_PARAMS);
		return;
	}
	/*
	 * We need to compute ceil(clk_freq * period / 500e6) - 2
	 * without incurring overflow.
	 */
	v = (ms->clk_freq / 5000) * period;
	if (v <= 250000) {
		/* special case: sync_period == 5 * clk_period */
		v = 0;
		/* units of tr are 100kB/s */
		tr = (ms->clk_freq + 250000) / 500000;
	} else {
		/* sync_period == (v + 2) * 2 * clk_period */
		v = (v + 99999) / 100000 - 2;
		if (v > 15)
			v = 15;	/* oops */
		tr = ((ms->clk_freq / (v + 2)) + 199999) / 200000;
	}
	if (offset > 15)
		offset = 15;	/* can't happen */
	tp->sync_params = SYNC_PARAMS(offset, v);
	out_8(&mr->sync_params, tp->sync_params);
	printk(KERN_INFO "mesh: target %d synchronous at %d.%d MB/s\n",
	       ms->conn_tgt, tr/10, tr%10);
}

static void start_phase(struct mesh_state *ms)
{
	int i, seq, nb;
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	volatile struct dbdma_regs __iomem *md = ms->dma;
	struct scsi_cmnd *cmd = ms->current_req;
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];

	dlog(ms, "start_phase nmo/exc/fc/seq = %.8x",
	     MKWORD(ms->n_msgout, mr->exception, mr->fifo_count, mr->sequence));
	out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	seq = use_active_neg + (ms->n_msgout? SEQ_ATN: 0);
	switch (ms->msgphase) {
	case msg_none:
		break;

	case msg_in:
		out_8(&mr->count_hi, 0);
		out_8(&mr->count_lo, 1);
		out_8(&mr->sequence, SEQ_MSGIN + seq);
		ms->n_msgin = 0;
		return;

	case msg_out:
		/*
		 * To make sure ATN drops before we assert ACK for
		 * the last byte of the message, we have to do the
		 * last byte specially.
		 */
		if (ms->n_msgout <= 0) {
			printk(KERN_ERR "mesh: msg_out but n_msgout=%d\n",
			       ms->n_msgout);
			mesh_dump_regs(ms);
			ms->msgphase = msg_none;
			break;
		}
		if (ALLOW_DEBUG(ms->conn_tgt)) {
			printk(KERN_DEBUG "mesh: sending %d msg bytes:",
			       ms->n_msgout);
			for (i = 0; i < ms->n_msgout; ++i)
				printk(" %x", ms->msgout[i]);
			printk("\n");
		}
		dlog(ms, "msgout msg=%.8x", MKWORD(ms->n_msgout, ms->msgout[0],
						ms->msgout[1], ms->msgout[2]));
		out_8(&mr->count_hi, 0);
		out_8(&mr->sequence, SEQ_FLUSHFIFO);
		mesh_flush_io(mr);
		udelay(1);
		/*
		 * If ATN is not already asserted, we assert it, then
		 * issue a SEQ_MSGOUT to get the mesh to drop ACK.
		 */
		if ((in_8(&mr->bus_status0) & BS0_ATN) == 0) {
			dlog(ms, "bus0 was %.2x explictly asserting ATN", mr->bus_status0);
			out_8(&mr->bus_status0, BS0_ATN); /* explicit ATN */
			mesh_flush_io(mr);
			udelay(1);
			out_8(&mr->count_lo, 1);
			out_8(&mr->sequence, SEQ_MSGOUT + seq);
			out_8(&mr->bus_status0, 0); /* release explicit ATN */
			dlog(ms,"hace: after explicit ATN bus0=%.2x",mr->bus_status0);
		}
		if (ms->n_msgout == 1) {
			/*
			 * We can't issue the SEQ_MSGOUT without ATN
			 * until the target has asserted REQ.  The logic
			 * in cmd_complete handles both situations:
			 * REQ already asserted or not.
			 */
			cmd_complete(ms);
		} else {
			out_8(&mr->count_lo, ms->n_msgout - 1);
			out_8(&mr->sequence, SEQ_MSGOUT + seq);
			for (i = 0; i < ms->n_msgout - 1; ++i)
				out_8(&mr->fifo, ms->msgout[i]);
		}
		return;

	default:
		printk(KERN_ERR "mesh bug: start_phase msgphase=%d\n",
		       ms->msgphase);
	}

	switch (ms->phase) {
	case selecting:
		out_8(&mr->dest_id, ms->conn_tgt);
		out_8(&mr->sequence, SEQ_SELECT + SEQ_ATN);
		break;
	case commanding:
		out_8(&mr->sync_params, tp->sync_params);
		out_8(&mr->count_hi, 0);
		if (cmd) {
			out_8(&mr->count_lo, cmd->cmd_len);
			out_8(&mr->sequence, SEQ_COMMAND + seq);
			for (i = 0; i < cmd->cmd_len; ++i)
				out_8(&mr->fifo, cmd->cmnd[i]);
		} else {
			out_8(&mr->count_lo, 6);
			out_8(&mr->sequence, SEQ_COMMAND + seq);
			for (i = 0; i < 6; ++i)
				out_8(&mr->fifo, 0);
		}
		break;
	case dataing:
		/* transfer data, if any */
		if (!ms->dma_started) {
			set_dma_cmds(ms, cmd);
			out_le32(&md->cmdptr, virt_to_phys(ms->dma_cmds));
			out_le32(&md->control, (RUN << 16) | RUN);
			ms->dma_started = 1;
		}
		nb = ms->dma_count;
		if (nb > 0xfff0)
			nb = 0xfff0;
		ms->dma_count -= nb;
		ms->data_ptr += nb;
		out_8(&mr->count_lo, nb);
		out_8(&mr->count_hi, nb >> 8);
		out_8(&mr->sequence, (tp->data_goes_out?
				SEQ_DATAOUT: SEQ_DATAIN) + SEQ_DMA_MODE + seq);
		break;
	case statusing:
		out_8(&mr->count_hi, 0);
		out_8(&mr->count_lo, 1);
		out_8(&mr->sequence, SEQ_STATUS + seq);
		break;
	case busfreeing:
	case disconnecting:
		out_8(&mr->sequence, SEQ_ENBRESEL);
		mesh_flush_io(mr);
		udelay(1);
		dlog(ms, "enbresel intr/exc/err/fc=%.8x",
		     MKWORD(mr->interrupt, mr->exception, mr->error,
			    mr->fifo_count));
		out_8(&mr->sequence, SEQ_BUSFREE);
		break;
	default:
		printk(KERN_ERR "mesh: start_phase called with phase=%d\n",
		       ms->phase);
		dumpslog(ms);
	}

}

static inline void get_msgin(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	int i, n;

	n = mr->fifo_count;
	if (n != 0) {
		i = ms->n_msgin;
		ms->n_msgin = i + n;
		for (; n > 0; --n)
			ms->msgin[i++] = in_8(&mr->fifo);
	}
}

static inline int msgin_length(struct mesh_state *ms)
{
	int b, n;

	n = 1;
	if (ms->n_msgin > 0) {
		b = ms->msgin[0];
		if (b == 1) {
			/* extended message */
			n = ms->n_msgin < 2? 2: ms->msgin[1] + 2;
		} else if (0x20 <= b && b <= 0x2f) {
			/* 2-byte message */
			n = 2;
		}
	}
	return n;
}

static void reselected(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	struct scsi_cmnd *cmd;
	struct mesh_target *tp;
	int b, t, prev;

	switch (ms->phase) {
	case idle:
		break;
	case arbitrating:
		if ((cmd = ms->current_req) != NULL) {
			/* put the command back on the queue */
			cmd->host_scribble = (void *) ms->request_q;
			if (ms->request_q == NULL)
				ms->request_qtail = cmd;
			ms->request_q = cmd;
			tp = &ms->tgts[cmd->device->id];
			tp->current_req = NULL;
		}
		break;
	case busfreeing:
		ms->phase = reselecting;
		mesh_done(ms, 0);
		break;
	case disconnecting:
		break;
	default:
		printk(KERN_ERR "mesh: reselected in phase %d/%d tgt %d\n",
		       ms->msgphase, ms->phase, ms->conn_tgt);
		dumplog(ms, ms->conn_tgt);
		dumpslog(ms);
	}

	if (ms->dma_started) {
		printk(KERN_ERR "mesh: reselected with DMA started !\n");
		halt_dma(ms);
	}
	ms->current_req = NULL;
	ms->phase = dataing;
	ms->msgphase = msg_in;
	ms->n_msgout = 0;
	ms->last_n_msgout = 0;
	prev = ms->conn_tgt;

	/*
	 * We seem to get abortive reselections sometimes.
	 */
	while ((in_8(&mr->bus_status1) & BS1_BSY) == 0) {
		static int mesh_aborted_resels;
		mesh_aborted_resels++;
		out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
		mesh_flush_io(mr);
		udelay(1);
		out_8(&mr->sequence, SEQ_ENBRESEL);
		mesh_flush_io(mr);
		udelay(5);
		dlog(ms, "extra resel err/exc/fc = %.6x",
		     MKWORD(0, mr->error, mr->exception, mr->fifo_count));
	}
	out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
       	mesh_flush_io(mr);
	udelay(1);
	out_8(&mr->sequence, SEQ_ENBRESEL);
       	mesh_flush_io(mr);
	udelay(1);
	out_8(&mr->sync_params, ASYNC_PARAMS);

	/*
	 * Find out who reselected us.
	 */
	if (in_8(&mr->fifo_count) == 0) {
		printk(KERN_ERR "mesh: reselection but nothing in fifo?\n");
		ms->conn_tgt = ms->host->this_id;
		goto bogus;
	}
	/* get the last byte in the fifo */
	do {
		b = in_8(&mr->fifo);
		dlog(ms, "reseldata %x", b);
	} while (in_8(&mr->fifo_count));
	for (t = 0; t < 8; ++t)
		if ((b & (1 << t)) != 0 && t != ms->host->this_id)
			break;
	if (b != (1 << t) + (1 << ms->host->this_id)) {
		printk(KERN_ERR "mesh: bad reselection data %x\n", b);
		ms->conn_tgt = ms->host->this_id;
		goto bogus;
	}


	/*
	 * Set up to continue with that target's transfer.
	 */
	ms->conn_tgt = t;
	tp = &ms->tgts[t];
	out_8(&mr->sync_params, tp->sync_params);
	if (ALLOW_DEBUG(t)) {
		printk(KERN_DEBUG "mesh: reselected by target %d\n", t);
		printk(KERN_DEBUG "mesh: saved_ptr=%x goes_out=%d cmd=%p\n",
		       tp->saved_ptr, tp->data_goes_out, tp->current_req);
	}
	ms->current_req = tp->current_req;
	if (tp->current_req == NULL) {
		printk(KERN_ERR "mesh: reselected by tgt %d but no cmd!\n", t);
		goto bogus;
	}
	ms->data_ptr = tp->saved_ptr;
	dlog(ms, "resel prev tgt=%d", prev);
	dlog(ms, "resel err/exc=%.4x", MKWORD(0, 0, mr->error, mr->exception));
	start_phase(ms);
	return;

bogus:
	dumplog(ms, ms->conn_tgt);
	dumpslog(ms);
	ms->data_ptr = 0;
	ms->aborting = 1;
	start_phase(ms);
}

static void do_abort(struct mesh_state *ms)
{
	ms->msgout[0] = ABORT;
	ms->n_msgout = 1;
	ms->aborting = 1;
	ms->stat = DID_ABORT;
	dlog(ms, "abort", 0);
}

static void handle_reset(struct mesh_state *ms)
{
	int tgt;
	struct mesh_target *tp;
	struct scsi_cmnd *cmd;
	volatile struct mesh_regs __iomem *mr = ms->mesh;

	for (tgt = 0; tgt < 8; ++tgt) {
		tp = &ms->tgts[tgt];
		if ((cmd = tp->current_req) != NULL) {
			cmd->result = DID_RESET << 16;
			tp->current_req = NULL;
			mesh_completed(ms, cmd);
		}
		ms->tgts[tgt].sdtr_state = do_sdtr;
		ms->tgts[tgt].sync_params = ASYNC_PARAMS;
	}
	ms->current_req = NULL;
	while ((cmd = ms->request_q) != NULL) {
		ms->request_q = (struct scsi_cmnd *) cmd->host_scribble;
		cmd->result = DID_RESET << 16;
		mesh_completed(ms, cmd);
	}
	ms->phase = idle;
	ms->msgphase = msg_none;
	out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	out_8(&mr->sequence, SEQ_FLUSHFIFO);
       	mesh_flush_io(mr);
	udelay(1);
	out_8(&mr->sync_params, ASYNC_PARAMS);
	out_8(&mr->sequence, SEQ_ENBRESEL);
}

static irqreturn_t do_mesh_interrupt(int irq, void *dev_id, struct pt_regs *ptregs)
{
	unsigned long flags;
	struct Scsi_Host *dev = ((struct mesh_state *)dev_id)->host;

	spin_lock_irqsave(dev->host_lock, flags);
	mesh_interrupt(irq, dev_id, ptregs);
	spin_unlock_irqrestore(dev->host_lock, flags);
	return IRQ_HANDLED;
}

static void handle_error(struct mesh_state *ms)
{
	int err, exc, count;
	volatile struct mesh_regs __iomem *mr = ms->mesh;

	err = in_8(&mr->error);
	exc = in_8(&mr->exception);
	out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	dlog(ms, "error err/exc/fc/cl=%.8x",
	     MKWORD(err, exc, mr->fifo_count, mr->count_lo));
	if (err & ERR_SCSIRESET) {
		/* SCSI bus was reset */
		printk(KERN_INFO "mesh: SCSI bus reset detected: "
		       "waiting for end...");
		while ((in_8(&mr->bus_status1) & BS1_RST) != 0)
			udelay(1);
		printk("done\n");
		handle_reset(ms);
		/* request_q is empty, no point in mesh_start() */
		return;
	}
	if (err & ERR_UNEXPDISC) {
		/* Unexpected disconnect */
		if (exc & EXC_RESELECTED) {
			reselected(ms);
			return;
		}
		if (!ms->aborting) {
			printk(KERN_WARNING "mesh: target %d aborted\n",
			       ms->conn_tgt);
			dumplog(ms, ms->conn_tgt);
			dumpslog(ms);
		}
		out_8(&mr->interrupt, INT_CMDDONE);
		ms->stat = DID_ABORT;
		mesh_done(ms, 1);
		return;
	}
	if (err & ERR_PARITY) {
		if (ms->msgphase == msg_in) {
			printk(KERN_ERR "mesh: msg parity error, target %d\n",
			       ms->conn_tgt);
			ms->msgout[0] = MSG_PARITY_ERROR;
			ms->n_msgout = 1;
			ms->msgphase = msg_in_bad;
			cmd_complete(ms);
			return;
		}
		if (ms->stat == DID_OK) {
			printk(KERN_ERR "mesh: parity error, target %d\n",
			       ms->conn_tgt);
			ms->stat = DID_PARITY;
		}
		count = (mr->count_hi << 8) + mr->count_lo;
		if (count == 0) {
			cmd_complete(ms);
		} else {
			/* reissue the data transfer command */
			out_8(&mr->sequence, mr->sequence);
		}
		return;
	}
	if (err & ERR_SEQERR) {
		if (exc & EXC_RESELECTED) {
			/* This can happen if we issue a command to
			   get the bus just after the target reselects us. */
			static int mesh_resel_seqerr;
			mesh_resel_seqerr++;
			reselected(ms);
			return;
		}
		if (exc == EXC_PHASEMM) {
			static int mesh_phasemm_seqerr;
			mesh_phasemm_seqerr++;
			phase_mismatch(ms);
			return;
		}
		printk(KERN_ERR "mesh: sequence error (err=%x exc=%x)\n",
		       err, exc);
	} else {
		printk(KERN_ERR "mesh: unknown error %x (exc=%x)\n", err, exc);
	}
	mesh_dump_regs(ms);
	dumplog(ms, ms->conn_tgt);
	if (ms->phase > selecting && (in_8(&mr->bus_status1) & BS1_BSY)) {
		/* try to do what the target wants */
		do_abort(ms);
		phase_mismatch(ms);
		return;
	}
	ms->stat = DID_ERROR;
	mesh_done(ms, 1);
}

static void handle_exception(struct mesh_state *ms)
{
	int exc;
	volatile struct mesh_regs __iomem *mr = ms->mesh;

	exc = in_8(&mr->exception);
	out_8(&mr->interrupt, INT_EXCEPTION | INT_CMDDONE);
	if (exc & EXC_RESELECTED) {
		static int mesh_resel_exc;
		mesh_resel_exc++;
		reselected(ms);
	} else if (exc == EXC_ARBLOST) {
		printk(KERN_DEBUG "mesh: lost arbitration\n");
		ms->stat = DID_BUS_BUSY;
		mesh_done(ms, 1);
	} else if (exc == EXC_SELTO) {
		/* selection timed out */
		ms->stat = DID_BAD_TARGET;
		mesh_done(ms, 1);
	} else if (exc == EXC_PHASEMM) {
		/* target wants to do something different:
		   find out what it wants and do it. */
		phase_mismatch(ms);
	} else {
		printk(KERN_ERR "mesh: can't cope with exception %x\n", exc);
		mesh_dump_regs(ms);
		dumplog(ms, ms->conn_tgt);
		do_abort(ms);
		phase_mismatch(ms);
	}
}

static void handle_msgin(struct mesh_state *ms)
{
	int i, code;
	struct scsi_cmnd *cmd = ms->current_req;
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];

	if (ms->n_msgin == 0)
		return;
	code = ms->msgin[0];
	if (ALLOW_DEBUG(ms->conn_tgt)) {
		printk(KERN_DEBUG "got %d message bytes:", ms->n_msgin);
		for (i = 0; i < ms->n_msgin; ++i)
			printk(" %x", ms->msgin[i]);
		printk("\n");
	}
	dlog(ms, "msgin msg=%.8x",
	     MKWORD(ms->n_msgin, code, ms->msgin[1], ms->msgin[2]));

	ms->expect_reply = 0;
	ms->n_msgout = 0;
	if (ms->n_msgin < msgin_length(ms))
		goto reject;
	if (cmd)
		cmd->SCp.Message = code;
	switch (code) {
	case COMMAND_COMPLETE:
		break;
	case EXTENDED_MESSAGE:
		switch (ms->msgin[2]) {
		case EXTENDED_MODIFY_DATA_POINTER:
			ms->data_ptr += (ms->msgin[3] << 24) + ms->msgin[6]
				+ (ms->msgin[4] << 16) + (ms->msgin[5] << 8);
			break;
		case EXTENDED_SDTR:
			if (tp->sdtr_state != sdtr_sent) {
				/* reply with an SDTR */
				add_sdtr_msg(ms);
				/* limit period to at least his value,
				   offset to no more than his */
				if (ms->msgout[3] < ms->msgin[3])
					ms->msgout[3] = ms->msgin[3];
				if (ms->msgout[4] > ms->msgin[4])
					ms->msgout[4] = ms->msgin[4];
				set_sdtr(ms, ms->msgout[3], ms->msgout[4]);
				ms->msgphase = msg_out;
			} else {
				set_sdtr(ms, ms->msgin[3], ms->msgin[4]);
			}
			break;
		default:
			goto reject;
		}
		break;
	case SAVE_POINTERS:
		tp->saved_ptr = ms->data_ptr;
		break;
	case RESTORE_POINTERS:
		ms->data_ptr = tp->saved_ptr;
		break;
	case DISCONNECT:
		ms->phase = disconnecting;
		break;
	case ABORT:
		break;
	case MESSAGE_REJECT:
		if (tp->sdtr_state == sdtr_sent)
			set_sdtr(ms, 0, 0);
		break;
	case NOP:
		break;
	default:
		if (IDENTIFY_BASE <= code && code <= IDENTIFY_BASE + 7) {
			if (cmd == NULL) {
				do_abort(ms);
				ms->msgphase = msg_out;
			} else if (code != cmd->device->lun + IDENTIFY_BASE) {
				printk(KERN_WARNING "mesh: lun mismatch "
				       "(%d != %d) on reselection from "
				       "target %d\n", code - IDENTIFY_BASE,
				       cmd->device->lun, ms->conn_tgt);
			}
			break;
		}
		goto reject;
	}
	return;

 reject:
	printk(KERN_WARNING "mesh: rejecting message from target %d:",
	       ms->conn_tgt);
	for (i = 0; i < ms->n_msgin; ++i)
		printk(" %x", ms->msgin[i]);
	printk("\n");
	ms->msgout[0] = MESSAGE_REJECT;
	ms->n_msgout = 1;
	ms->msgphase = msg_out;
}

/*
 * Set up DMA commands for transferring data.
 */
static void set_dma_cmds(struct mesh_state *ms, struct scsi_cmnd *cmd)
{
	int i, dma_cmd, total, off, dtot;
	struct scatterlist *scl;
	struct dbdma_cmd *dcmds;

	dma_cmd = ms->tgts[ms->conn_tgt].data_goes_out?
		OUTPUT_MORE: INPUT_MORE;
	dcmds = ms->dma_cmds;
	dtot = 0;
	if (cmd) {
		cmd->SCp.this_residual = cmd->request_bufflen;
		if (cmd->use_sg > 0) {
			int nseg;
			total = 0;
			scl = (struct scatterlist *) cmd->buffer;
			off = ms->data_ptr;
			nseg = pci_map_sg(ms->pdev, scl, cmd->use_sg,
					  cmd->sc_data_direction);
			for (i = 0; i <nseg; ++i, ++scl) {
				u32 dma_addr = sg_dma_address(scl);
				u32 dma_len = sg_dma_len(scl);

				total += scl->length;
				if (off >= dma_len) {
					off -= dma_len;
					continue;
				}
				if (dma_len > 0xffff)
					panic("mesh: scatterlist element >= 64k");
				st_le16(&dcmds->req_count, dma_len - off);
				st_le16(&dcmds->command, dma_cmd);
				st_le32(&dcmds->phy_addr, dma_addr + off);
				dcmds->xfer_status = 0;
				++dcmds;
				dtot += dma_len - off;
				off = 0;
			}
		} else if (ms->data_ptr < cmd->request_bufflen) {
			dtot = cmd->request_bufflen - ms->data_ptr;
			if (dtot > 0xffff)
				panic("mesh: transfer size >= 64k");
			st_le16(&dcmds->req_count, dtot);
			/* XXX Use pci DMA API here ... */
			st_le32(&dcmds->phy_addr,
				virt_to_phys(cmd->request_buffer) + ms->data_ptr);
			dcmds->xfer_status = 0;
			++dcmds;
		}
	}
	if (dtot == 0) {
		/* Either the target has overrun our buffer,
		   or the caller didn't provide a buffer. */
		static char mesh_extra_buf[64];

		dtot = sizeof(mesh_extra_buf);
		st_le16(&dcmds->req_count, dtot);
		st_le32(&dcmds->phy_addr, virt_to_phys(mesh_extra_buf));
		dcmds->xfer_status = 0;
		++dcmds;
	}
	dma_cmd += OUTPUT_LAST - OUTPUT_MORE;
	st_le16(&dcmds[-1].command, dma_cmd);
	memset(dcmds, 0, sizeof(*dcmds));
	st_le16(&dcmds->command, DBDMA_STOP);
	ms->dma_count = dtot;
}

static void halt_dma(struct mesh_state *ms)
{
	volatile struct dbdma_regs __iomem *md = ms->dma;
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	struct scsi_cmnd *cmd = ms->current_req;
	int t, nb;

	if (!ms->tgts[ms->conn_tgt].data_goes_out) {
		/* wait a little while until the fifo drains */
		t = 50;
		while (t > 0 && in_8(&mr->fifo_count) != 0
		       && (in_le32(&md->status) & ACTIVE) != 0) {
			--t;
			udelay(1);
		}
	}
	out_le32(&md->control, RUN << 16);	/* turn off RUN bit */
	nb = (mr->count_hi << 8) + mr->count_lo;
	dlog(ms, "halt_dma fc/count=%.6x",
	     MKWORD(0, mr->fifo_count, 0, nb));
	if (ms->tgts[ms->conn_tgt].data_goes_out)
		nb += mr->fifo_count;
	/* nb is the number of bytes not yet transferred
	   to/from the target. */
	ms->data_ptr -= nb;
	dlog(ms, "data_ptr %x", ms->data_ptr);
	if (ms->data_ptr < 0) {
		printk(KERN_ERR "mesh: halt_dma: data_ptr=%d (nb=%d, ms=%p)\n",
		       ms->data_ptr, nb, ms);
		ms->data_ptr = 0;
#ifdef MESH_DBG
		dumplog(ms, ms->conn_tgt);
		dumpslog(ms);
#endif /* MESH_DBG */
	} else if (cmd && cmd->request_bufflen != 0 &&
		   ms->data_ptr > cmd->request_bufflen) {
		printk(KERN_DEBUG "mesh: target %d overrun, "
		       "data_ptr=%x total=%x goes_out=%d\n",
		       ms->conn_tgt, ms->data_ptr, cmd->request_bufflen,
		       ms->tgts[ms->conn_tgt].data_goes_out);
	}
	if (cmd->use_sg != 0) {
		struct scatterlist *sg;
		sg = (struct scatterlist *)cmd->request_buffer;
		pci_unmap_sg(ms->pdev, sg, cmd->use_sg, cmd->sc_data_direction);
	}
	ms->dma_started = 0;
}

static void phase_mismatch(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	int phase;

	dlog(ms, "phasemm ch/cl/seq/fc=%.8x",
	     MKWORD(mr->count_hi, mr->count_lo, mr->sequence, mr->fifo_count));
	phase = in_8(&mr->bus_status0) & BS0_PHASE;
	if (ms->msgphase == msg_out_xxx && phase == BP_MSGOUT) {
		/* output the last byte of the message, without ATN */
		out_8(&mr->count_lo, 1);
		out_8(&mr->sequence, SEQ_MSGOUT + use_active_neg);
		mesh_flush_io(mr);
		udelay(1);
		out_8(&mr->fifo, ms->msgout[ms->n_msgout-1]);
		ms->msgphase = msg_out_last;
		return;
	}

	if (ms->msgphase == msg_in) {
		get_msgin(ms);
		if (ms->n_msgin)
			handle_msgin(ms);
	}

	if (ms->dma_started)
		halt_dma(ms);
	if (mr->fifo_count) {
		out_8(&mr->sequence, SEQ_FLUSHFIFO);
		mesh_flush_io(mr);
		udelay(1);
	}

	ms->msgphase = msg_none;
	switch (phase) {
	case BP_DATAIN:
		ms->tgts[ms->conn_tgt].data_goes_out = 0;
		ms->phase = dataing;
		break;
	case BP_DATAOUT:
		ms->tgts[ms->conn_tgt].data_goes_out = 1;
		ms->phase = dataing;
		break;
	case BP_COMMAND:
		ms->phase = commanding;
		break;
	case BP_STATUS:
		ms->phase = statusing;
		break;
	case BP_MSGIN:
		ms->msgphase = msg_in;
		ms->n_msgin = 0;
		break;
	case BP_MSGOUT:
		ms->msgphase = msg_out;
		if (ms->n_msgout == 0) {
			if (ms->aborting) {
				do_abort(ms);
			} else {
				if (ms->last_n_msgout == 0) {
					printk(KERN_DEBUG
					       "mesh: no msg to repeat\n");
					ms->msgout[0] = NOP;
					ms->last_n_msgout = 1;
				}
				ms->n_msgout = ms->last_n_msgout;
			}
		}
		break;
	default:
		printk(KERN_DEBUG "mesh: unknown scsi phase %x\n", phase);
		ms->stat = DID_ERROR;
		mesh_done(ms, 1);
		return;
	}

	start_phase(ms);
}

static void cmd_complete(struct mesh_state *ms)
{
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	struct scsi_cmnd *cmd = ms->current_req;
	struct mesh_target *tp = &ms->tgts[ms->conn_tgt];
	int seq, n, t;

	dlog(ms, "cmd_complete fc=%x", mr->fifo_count);
	seq = use_active_neg + (ms->n_msgout? SEQ_ATN: 0);
	switch (ms->msgphase) {
	case msg_out_xxx:
		/* huh?  we expected a phase mismatch */
		ms->n_msgin = 0;
		ms->msgphase = msg_in;
		/* fall through */

	case msg_in:
		/* should have some message bytes in fifo */
		get_msgin(ms);
		n = msgin_length(ms);
		if (ms->n_msgin < n) {
			out_8(&mr->count_lo, n - ms->n_msgin);
			out_8(&mr->sequence, SEQ_MSGIN + seq);
		} else {
			ms->msgphase = msg_none;
			handle_msgin(ms);
			start_phase(ms);
		}
		break;

	case msg_in_bad:
		out_8(&mr->sequence, SEQ_FLUSHFIFO);
		mesh_flush_io(mr);
		udelay(1);
		out_8(&mr->count_lo, 1);
		out_8(&mr->sequence, SEQ_MSGIN + SEQ_ATN + use_active_neg);
		break;

	case msg_out:
		/*
		 * To get the right timing on ATN wrt ACK, we have
		 * to get the MESH to drop ACK, wait until REQ gets
		 * asserted, then drop ATN.  To do this we first
		 * issue a SEQ_MSGOUT with ATN and wait for REQ,
		 * then change the command to a SEQ_MSGOUT w/o ATN.
		 * If we don't see REQ in a reasonable time, we
		 * change the command to SEQ_MSGIN with ATN,
		 * wait for the phase mismatch interrupt, then
		 * issue the SEQ_MSGOUT without ATN.
		 */
		out_8(&mr->count_lo, 1);
		out_8(&mr->sequence, SEQ_MSGOUT + use_active_neg + SEQ_ATN);
		t = 30;		/* wait up to 30us */
		while ((in_8(&mr->bus_status0) & BS0_REQ) == 0 && --t >= 0)
			udelay(1);
		dlog(ms, "last_mbyte err/exc/fc/cl=%.8x",
		     MKWORD(mr->error, mr->exception,
			    mr->fifo_count, mr->count_lo));
		if (in_8(&mr->interrupt) & (INT_ERROR | INT_EXCEPTION)) {
			/* whoops, target didn't do what we expected */
			ms->last_n_msgout = ms->n_msgout;
			ms->n_msgout = 0;
			if (in_8(&mr->interrupt) & INT_ERROR) {
				printk(KERN_ERR "mesh: error %x in msg_out\n",
				       in_8(&mr->error));
				handle_error(ms);
				return;
			}
			if (in_8(&mr->exception) != EXC_PHASEMM)
				printk(KERN_ERR "mesh: exc %x in msg_out\n",
				       in_8(&mr->exception));
			else
				printk(KERN_DEBUG "mesh: bs0=%x in msg_out\n",
				       in_8(&mr->bus_status0));
			handle_exception(ms);
			return;
		}
		if (in_8(&mr->bus_status0) & BS0_REQ) {
			out_8(&mr->sequence, SEQ_MSGOUT + use_active_neg);
			mesh_flush_io(mr);
			udelay(1);
			out_8(&mr->fifo, ms->msgout[ms->n_msgout-1]);
			ms->msgphase = msg_out_last;
		} else {
			out_8(&mr->sequence, SEQ_MSGIN + use_active_neg + SEQ_ATN);
			ms->msgphase = msg_out_xxx;
		}
		break;

	case msg_out_last:
		ms->last_n_msgout = ms->n_msgout;
		ms->n_msgout = 0;
		ms->msgphase = ms->expect_reply? msg_in: msg_none;
		start_phase(ms);
		break;

	case msg_none:
		switch (ms->phase) {
		case idle:
			printk(KERN_ERR "mesh: interrupt in idle phase?\n");
			dumpslog(ms);
			return;
		case selecting:
			dlog(ms, "Selecting phase at command completion",0);
			ms->msgout[0] = IDENTIFY(ALLOW_RESEL(ms->conn_tgt),
						 (cmd? cmd->device->lun: 0));
			ms->n_msgout = 1;
			ms->expect_reply = 0;
			if (ms->aborting) {
				ms->msgout[0] = ABORT;
				ms->n_msgout++;
			} else if (tp->sdtr_state == do_sdtr) {
				/* add SDTR message */
				add_sdtr_msg(ms);
				ms->expect_reply = 1;
				tp->sdtr_state = sdtr_sent;
			}
			ms->msgphase = msg_out;
			/*
			 * We need to wait for REQ before dropping ATN.
			 * We wait for at most 30us, then fall back to
			 * a scheme where we issue a SEQ_COMMAND with ATN,
			 * which will give us a phase mismatch interrupt
			 * when REQ does come, and then we send the message.
			 */
			t = 230;		/* wait up to 230us */
			while ((in_8(&mr->bus_status0) & BS0_REQ) == 0) {
				if (--t < 0) {
					dlog(ms, "impatient for req", ms->n_msgout);
					ms->msgphase = msg_none;
					break;
				}
				udelay(1);
			}
			break;
		case dataing:
			if (ms->dma_count != 0) {
				start_phase(ms);
				return;
			}
			/*
			 * We can get a phase mismatch here if the target
			 * changes to the status phase, even though we have
			 * had a command complete interrupt.  Then, if we
			 * issue the SEQ_STATUS command, we'll get a sequence
			 * error interrupt.  Which isn't so bad except that
			 * occasionally the mesh actually executes the
			 * SEQ_STATUS *as well as* giving us the sequence
			 * error and phase mismatch exception.
			 */
			out_8(&mr->sequence, 0);
			out_8(&mr->interrupt,
			      INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
			halt_dma(ms);
			break;
		case statusing:
			if (cmd) {
				cmd->SCp.Status = mr->fifo;
				if (DEBUG_TARGET(cmd))
					printk(KERN_DEBUG "mesh: status is %x\n",
					       cmd->SCp.Status);
			}
			ms->msgphase = msg_in;
			break;
		case busfreeing:
			mesh_done(ms, 1);
			return;
		case disconnecting:
			ms->current_req = NULL;
			ms->phase = idle;
			mesh_start(ms);
			return;
		default:
			break;
		}
		++ms->phase;
		start_phase(ms);
		break;
	}
}


/*
 * Called by midlayer with host locked to queue a new
 * request
 */
static int mesh_queue(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
{
	struct mesh_state *ms;

	cmd->scsi_done = done;
	cmd->host_scribble = NULL;

	ms = (struct mesh_state *) cmd->device->host->hostdata;

	if (ms->request_q == NULL)
		ms->request_q = cmd;
	else
		ms->request_qtail->host_scribble = (void *) cmd;
	ms->request_qtail = cmd;

	if (ms->phase == idle)
		mesh_start(ms);

	return 0;
}

/*
 * Called to handle interrupts, either call by the interrupt
 * handler (do_mesh_interrupt) or by other functions in
 * exceptional circumstances
 */
static void mesh_interrupt(int irq, void *dev_id, struct pt_regs *ptregs)
{
	struct mesh_state *ms = (struct mesh_state *) dev_id;
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	int intr;

#if 0
	if (ALLOW_DEBUG(ms->conn_tgt))
		printk(KERN_DEBUG "mesh_intr, bs0=%x int=%x exc=%x err=%x "
		       "phase=%d msgphase=%d\n", mr->bus_status0,
		       mr->interrupt, mr->exception, mr->error,
		       ms->phase, ms->msgphase);
#endif
	while ((intr = in_8(&mr->interrupt)) != 0) {
		dlog(ms, "interrupt intr/err/exc/seq=%.8x",
		     MKWORD(intr, mr->error, mr->exception, mr->sequence));
		if (intr & INT_ERROR) {
			handle_error(ms);
		} else if (intr & INT_EXCEPTION) {
			handle_exception(ms);
		} else if (intr & INT_CMDDONE) {
			out_8(&mr->interrupt, INT_CMDDONE);
			cmd_complete(ms);
		}
	}
}

/* Todo: here we can at least try to remove the command from the
 * queue if it isn't connected yet, and for pending command, assert
 * ATN until the bus gets freed.
 */
static int mesh_abort(struct scsi_cmnd *cmd)
{
	struct mesh_state *ms = (struct mesh_state *) cmd->device->host->hostdata;

	printk(KERN_DEBUG "mesh_abort(%p)\n", cmd);
	mesh_dump_regs(ms);
	dumplog(ms, cmd->device->id);
	dumpslog(ms);
	return FAILED;
}

/*
 * Called by the midlayer with the lock held to reset the
 * SCSI host and bus.
 * The midlayer will wait for devices to come back, we don't need
 * to do that ourselves
 */
static int mesh_host_reset(struct scsi_cmnd *cmd)
{
	struct mesh_state *ms = (struct mesh_state *) cmd->device->host->hostdata;
	volatile struct mesh_regs __iomem *mr = ms->mesh;
	volatile struct dbdma_regs __iomem *md = ms->dma;

	printk(KERN_DEBUG "mesh_host_reset\n");

	/* Reset the controller & dbdma channel */
	out_le32(&md->control, (RUN|PAUSE|FLUSH|WAKE) << 16);	/* stop dma */
	out_8(&mr->exception, 0xff);	/* clear all exception bits */
	out_8(&mr->error, 0xff);	/* clear all error bits */
	out_8(&mr->sequence, SEQ_RESETMESH);
       	mesh_flush_io(mr);
	udelay(1);
	out_8(&mr->intr_mask, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	out_8(&mr->source_id, ms->host->this_id);
	out_8(&mr->sel_timeout, 25);	/* 250ms */
	out_8(&mr->sync_params, ASYNC_PARAMS);

	/* Reset the bus */
	out_8(&mr->bus_status1, BS1_RST);	/* assert RST */
       	mesh_flush_io(mr);
	udelay(30);			/* leave it on for >= 25us */
	out_8(&mr->bus_status1, 0);	/* negate RST */

	/* Complete pending commands */
	handle_reset(ms);

	return SUCCESS;
}

static void set_mesh_power(struct mesh_state *ms, int state)
{
	if (_machine != _MACH_Pmac)
		return;
	if (state) {
		pmac_call_feature(PMAC_FTR_MESH_ENABLE, macio_get_of_node(ms->mdev), 0, 1);
		msleep(200);
	} else {
		pmac_call_feature(PMAC_FTR_MESH_ENABLE, macio_get_of_node(ms->mdev), 0, 0);
		msleep(10);
	}
}


#ifdef CONFIG_PM
static int mesh_suspend(struct macio_dev *mdev, u32 state)
{
	struct mesh_state *ms = (struct mesh_state *)macio_get_drvdata(mdev);
	unsigned long flags;

	if (state == mdev->ofdev.dev.power.power_state || state < 2)
		return 0;

	scsi_block_requests(ms->host);
	spin_lock_irqsave(ms->host->host_lock, flags);
	while(ms->phase != idle) {
		spin_unlock_irqrestore(ms->host->host_lock, flags);
		msleep(10);
		spin_lock_irqsave(ms->host->host_lock, flags);
	}
	ms->phase = sleeping;
	spin_unlock_irqrestore(ms->host->host_lock, flags);
	disable_irq(ms->meshintr);
	set_mesh_power(ms, 0);

	mdev->ofdev.dev.power.power_state = state;

	return 0;
}

static int mesh_resume(struct macio_dev *mdev)
{
	struct mesh_state *ms = (struct mesh_state *)macio_get_drvdata(mdev);
	unsigned long flags;

	if (mdev->ofdev.dev.power.power_state == 0)
		return 0;

	set_mesh_power(ms, 1);
	mesh_init(ms);
	spin_lock_irqsave(ms->host->host_lock, flags);
	mesh_start(ms);
	spin_unlock_irqrestore(ms->host->host_lock, flags);
	enable_irq(ms->meshintr);
	scsi_unblock_requests(ms->host);

	mdev->ofdev.dev.power.power_state = 0;

	return 0;
}

#endif /* CONFIG_PM */

/*
 * If we leave drives set for synchronous transfers (especially
 * CDROMs), and reboot to MacOS, it gets confused, poor thing.
 * So, on reboot we reset the SCSI bus.
 */
static int mesh_shutdown(struct macio_dev *mdev)
{
	struct mesh_state *ms = (struct mesh_state *)macio_get_drvdata(mdev);
	volatile struct mesh_regs __iomem *mr;
	unsigned long flags;

       	printk(KERN_INFO "resetting MESH scsi bus(es)\n");
	spin_lock_irqsave(ms->host->host_lock, flags);
       	mr = ms->mesh;
	out_8(&mr->intr_mask, 0);
	out_8(&mr->interrupt, INT_ERROR | INT_EXCEPTION | INT_CMDDONE);
	out_8(&mr->bus_status1, BS1_RST);
	mesh_flush_io(mr);
	udelay(30);
	out_8(&mr->bus_status1, 0);
	spin_unlock_irqrestore(ms->host->host_lock, flags);

	return 0;
}

static struct scsi_host_template mesh_template = {
	.proc_name			= "mesh",
	.name				= "MESH",
	.queuecommand			= mesh_queue,
	.eh_abort_handler		= mesh_abort,
	.eh_host_reset_handler		= mesh_host_reset,
	.can_queue			= 20,
	.this_id			= 7,
	.sg_tablesize			= SG_ALL,
	.cmd_per_lun			= 2,
	.use_clustering			= DISABLE_CLUSTERING,
};

static int mesh_probe(struct macio_dev *mdev, const struct of_match *match)
{
	struct device_node *mesh = macio_get_of_node(mdev);
	struct pci_dev* pdev = macio_get_pci_dev(mdev);
	int tgt, *cfp, minper;
	struct mesh_state *ms;
	struct Scsi_Host *mesh_host;
	void *dma_cmd_space;
	dma_addr_t dma_cmd_bus;

	switch (mdev->bus->chip->type) {
	case macio_heathrow:
	case macio_gatwick:
	case macio_paddington:
		use_active_neg = 0;
		break;
	default:
		use_active_neg = SEQ_ACTIVE_NEG;
	}

	if (macio_resource_count(mdev) != 2 || macio_irq_count(mdev) != 2) {
       		printk(KERN_ERR "mesh: expected 2 addrs and 2 intrs"
	       	       " (got %d,%d)\n", mesh->n_addrs, mesh->n_intrs);
		return -ENODEV;
	}

	if (macio_request_resources(mdev, "mesh") != 0) {
       		printk(KERN_ERR "mesh: unable to request memory resources");
		return -EBUSY;
	}
       	mesh_host = scsi_host_alloc(&mesh_template, sizeof(struct mesh_state));
	if (mesh_host == NULL) {
		printk(KERN_ERR "mesh: couldn't register host");
		goto out_release;
	}

	/* Old junk for root discovery, that will die ultimately */
#if !defined(MODULE)
       	note_scsi_host(mesh, mesh_host);
#endif

	mesh_host->base = macio_resource_start(mdev, 0);
	mesh_host->irq = macio_irq(mdev, 0);
       	ms = (struct mesh_state *) mesh_host->hostdata;
	macio_set_drvdata(mdev, ms);
	ms->host = mesh_host;
	ms->mdev = mdev;
	ms->pdev = pdev;

	ms->mesh = ioremap(macio_resource_start(mdev, 0), 0x1000);
	if (ms->mesh == NULL) {
		printk(KERN_ERR "mesh: can't map registers\n");
		goto out_free;
	}
	ms->dma = ioremap(macio_resource_start(mdev, 1), 0x1000);
	if (ms->dma == NULL) {
		printk(KERN_ERR "mesh: can't map registers\n");
		iounmap(ms->mesh);
		goto out_free;
	}

       	ms->meshintr = macio_irq(mdev, 0);
       	ms->dmaintr = macio_irq(mdev, 1);

       	/* Space for dma command list: +1 for stop command,
       	 * +1 to allow for aligning.
	 */
	ms->dma_cmd_size = (mesh_host->sg_tablesize + 2) * sizeof(struct dbdma_cmd);

	/* We use the PCI APIs for now until the generic one gets fixed
	 * enough or until we get some macio-specific versions
	 */
	dma_cmd_space = pci_alloc_consistent(macio_get_pci_dev(mdev),
					     ms->dma_cmd_size,
					     &dma_cmd_bus);
	if (dma_cmd_space == NULL) {
		printk(KERN_ERR "mesh: can't allocate DMA table\n");
		goto out_unmap;
	}
	memset(dma_cmd_space, 0, ms->dma_cmd_size);

	ms->dma_cmds = (struct dbdma_cmd *) DBDMA_ALIGN(dma_cmd_space);
       	ms->dma_cmd_space = dma_cmd_space;
	ms->dma_cmd_bus = dma_cmd_bus + ((unsigned long)ms->dma_cmds)
		- (unsigned long)dma_cmd_space;
	ms->current_req = NULL;
       	for (tgt = 0; tgt < 8; ++tgt) {
	       	ms->tgts[tgt].sdtr_state = do_sdtr;
	       	ms->tgts[tgt].sync_params = ASYNC_PARAMS;
	       	ms->tgts[tgt].current_req = NULL;
       	}

	if ((cfp = (int *) get_property(mesh, "clock-frequency", NULL)))
       		ms->clk_freq = *cfp;
	else {
       		printk(KERN_INFO "mesh: assuming 50MHz clock frequency\n");
	       	ms->clk_freq = 50000000;
       	}

       	/* The maximum sync rate is clock / 5; increase
       	 * mesh_sync_period if necessary.
	 */
	minper = 1000000000 / (ms->clk_freq / 5); /* ns */
	if (mesh_sync_period < minper)
		mesh_sync_period = minper;

	/* Power up the chip */
	set_mesh_power(ms, 1);

	/* Set it up */
       	mesh_init(ms);

	/* XXX FIXME: error should be fatal */
       	if (request_irq(ms->meshintr, do_mesh_interrupt, 0, "MESH", ms))
	       	printk(KERN_ERR "MESH: can't get irq %d\n", ms->meshintr);

	/* XXX FIXME: handle failure */
	scsi_add_host(mesh_host, &mdev->ofdev.dev);
	scsi_scan_host(mesh_host);

	return 0;

out_unmap:
	iounmap(ms->dma);
	iounmap(ms->mesh);
out_free:
	scsi_host_put(mesh_host);
out_release:
	macio_release_resources(mdev);

	return -ENODEV;
}

static int mesh_remove(struct macio_dev *mdev)
{
	struct mesh_state *ms = (struct mesh_state *)macio_get_drvdata(mdev);
	struct Scsi_Host *mesh_host = ms->host;

	scsi_remove_host(mesh_host);

	free_irq(ms->meshintr, ms);

	/* Reset scsi bus */
	mesh_shutdown(mdev);

	/* Shut down chip & termination */
	set_mesh_power(ms, 0);

	/* Unmap registers & dma controller */
	iounmap(ms->mesh);
       	iounmap(ms->dma);

	/* Free DMA commands memory */
	pci_free_consistent(macio_get_pci_dev(mdev), ms->dma_cmd_size,
			  ms->dma_cmd_space, ms->dma_cmd_bus);

	/* Release memory resources */
	macio_release_resources(mdev);

	scsi_host_put(mesh_host);

	return 0;
}


static struct of_match mesh_match[] =
{
	{
	.name 		= "mesh",
	.type		= OF_ANY_MATCH,
	.compatible	= OF_ANY_MATCH
	},
	{
	.name 		= OF_ANY_MATCH,
	.type		= "scsi",
	.compatible	= "chrp,mesh0"
	},
	{},
};

static struct macio_driver mesh_driver =
{
	.name 		= "mesh",
	.match_table	= mesh_match,
	.probe		= mesh_probe,
	.remove		= mesh_remove,
	.shutdown	= mesh_shutdown,
#ifdef CONFIG_PM
	.suspend	= mesh_suspend,
	.resume		= mesh_resume,
#endif
};


static int __init init_mesh(void)
{

	/* Calculate sync rate from module parameters */
	if (sync_rate > 10)
		sync_rate = 10;
	if (sync_rate > 0) {
		printk(KERN_INFO "mesh: configured for synchronous %d MB/s\n", sync_rate);
		mesh_sync_period = 1000 / sync_rate;	/* ns */
		mesh_sync_offset = 15;
	} else
		printk(KERN_INFO "mesh: configured for asynchronous\n");

	return macio_register_driver(&mesh_driver);
}

static void __exit exit_mesh(void)
{
	return macio_unregister_driver(&mesh_driver);
}

module_init(init_mesh);
module_exit(exit_mesh);