lba_pci.c

/*
**
**  PCI Lower Bus Adapter (LBA) manager
**
**	(c) Copyright 1999,2000 Grant Grundler
**	(c) Copyright 1999,2000 Hewlett-Packard Company
**
**	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.
**
**
** This module primarily provides access to PCI bus (config/IOport
** spaces) on platforms with an SBA/LBA chipset. A/B/C/J/L/N-class
** with 4 digit model numbers - eg C3000 (and A400...sigh).
**
** LBA driver isn't as simple as the Dino driver because:
**   (a) this chip has substantial bug fixes between revisions
**       (Only one Dino bug has a software workaround :^(  )
**   (b) has more options which we don't (yet) support (DMA hints, OLARD)
**   (c) IRQ support lives in the I/O SAPIC driver (not with PCI driver)
**   (d) play nicely with both PAT and "Legacy" PA-RISC firmware (PDC).
**       (dino only deals with "Legacy" PDC)
**
** LBA driver passes the I/O SAPIC HPA to the I/O SAPIC driver.
** (I/O SAPIC is integratd in the LBA chip).
**
** FIXME: Add support to SBA and LBA drivers for DMA hint sets
** FIXME: Add support for PCI card hot-plug (OLARD).
*/

#include <linux/delay.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/spinlock.h>
#include <linux/init.h>		/* for __init and __devinit */
#include <linux/pci.h>
#include <linux/ioport.h>
#include <linux/slab.h>

#include <asm/byteorder.h>
#include <asm/pdc.h>
#include <asm/pdcpat.h>
#include <asm/page.h>
#include <asm/system.h>

#include <asm/ropes.h>
#include <asm/hardware.h>	/* for register_parisc_driver() stuff */
#include <asm/parisc-device.h>
#include <asm/io.h>		/* read/write stuff */

#undef DEBUG_LBA	/* general stuff */
#undef DEBUG_LBA_PORT	/* debug I/O Port access */
#undef DEBUG_LBA_CFG	/* debug Config Space Access (ie PCI Bus walk) */
#undef DEBUG_LBA_PAT	/* debug PCI Resource Mgt code - PDC PAT only */

#undef FBB_SUPPORT	/* Fast Back-Back xfers - NOT READY YET */


#ifdef DEBUG_LBA
#define DBG(x...)	printk(x)
#else
#define DBG(x...)
#endif

#ifdef DEBUG_LBA_PORT
#define DBG_PORT(x...)	printk(x)
#else
#define DBG_PORT(x...)
#endif

#ifdef DEBUG_LBA_CFG
#define DBG_CFG(x...)	printk(x)
#else
#define DBG_CFG(x...)
#endif

#ifdef DEBUG_LBA_PAT
#define DBG_PAT(x...)	printk(x)
#else
#define DBG_PAT(x...)
#endif


/*
** Config accessor functions only pass in the 8-bit bus number and not
** the 8-bit "PCI Segment" number. Each LBA will be assigned a PCI bus
** number based on what firmware wrote into the scratch register.
**
** The "secondary" bus number is set to this before calling
** pci_register_ops(). If any PPB's are present, the scan will
** discover them and update the "secondary" and "subordinate"
** fields in the pci_bus structure.
**
** Changes in the configuration *may* result in a different
** bus number for each LBA depending on what firmware does.
*/

#define MODULE_NAME "LBA"

/* non-postable I/O port space, densely packed */
#define LBA_PORT_BASE	(PCI_F_EXTEND | 0xfee00000UL)
static void __iomem *astro_iop_base __read_mostly;

static u32 lba_t32;

/* lba flags */
#define LBA_FLAG_SKIP_PROBE	0x10

#define LBA_SKIP_PROBE(d) ((d)->flags & LBA_FLAG_SKIP_PROBE)


/* Looks nice and keeps the compiler happy */
#define LBA_DEV(d) ((struct lba_device *) (d))


/*
** Only allow 8 subsidiary busses per LBA
** Problem is the PCI bus numbering is globally shared.
*/
#define LBA_MAX_NUM_BUSES 8

/************************************
 * LBA register read and write support
 *
 * BE WARNED: register writes are posted.
 *  (ie follow writes which must reach HW with a read)
 */
#define READ_U8(addr)  __raw_readb(addr)
#define READ_U16(addr) __raw_readw(addr)
#define READ_U32(addr) __raw_readl(addr)
#define WRITE_U8(value, addr)  __raw_writeb(value, addr)
#define WRITE_U16(value, addr) __raw_writew(value, addr)
#define WRITE_U32(value, addr) __raw_writel(value, addr)

#define READ_REG8(addr)  readb(addr)
#define READ_REG16(addr) readw(addr)
#define READ_REG32(addr) readl(addr)
#define READ_REG64(addr) readq(addr)
#define WRITE_REG8(value, addr)  writeb(value, addr)
#define WRITE_REG16(value, addr) writew(value, addr)
#define WRITE_REG32(value, addr) writel(value, addr)


#define LBA_CFG_TOK(bus,dfn) ((u32) ((bus)<<16 | (dfn)<<8))
#define LBA_CFG_BUS(tok)  ((u8) ((tok)>>16))
#define LBA_CFG_DEV(tok)  ((u8) ((tok)>>11) & 0x1f)
#define LBA_CFG_FUNC(tok) ((u8) ((tok)>>8 ) & 0x7)


/*
** Extract LBA (Rope) number from HPA
** REVISIT: 16 ropes for Stretch/Ike?
*/
#define ROPES_PER_IOC	8
#define LBA_NUM(x)    ((((unsigned long) x) >> 13) & (ROPES_PER_IOC-1))


static void
lba_dump_res(struct resource *r, int d)
{
	int i;

	if (NULL == r)
		return;

	printk(KERN_DEBUG "(%p)", r->parent);
	for (i = d; i ; --i) printk(" ");
	printk(KERN_DEBUG "%p [%lx,%lx]/%lx\n", r,
		(long)r->start, (long)r->end, r->flags);
	lba_dump_res(r->child, d+2);
	lba_dump_res(r->sibling, d);
}


/*
** LBA rev 2.0, 2.1, 2.2, and 3.0 bus walks require a complex
** workaround for cfg cycles:
**	-- preserve  LBA state
**	-- prevent any DMA from occurring
**	-- turn on smart mode
**	-- probe with config writes before doing config reads
**	-- check ERROR_STATUS
**	-- clear ERROR_STATUS
**	-- restore LBA state
**
** The workaround is only used for device discovery.
*/

static int lba_device_present(u8 bus, u8 dfn, struct lba_device *d)
{
	u8 first_bus = d->hba.hba_bus->secondary;
	u8 last_sub_bus = d->hba.hba_bus->subordinate;

	if ((bus < first_bus) ||
	    (bus > last_sub_bus) ||
	    ((bus - first_bus) >= LBA_MAX_NUM_BUSES)) {
		return 0;
	}

	return 1;
}



#define LBA_CFG_SETUP(d, tok) {				\
    /* Save contents of error config register.  */			\
    error_config = READ_REG32(d->hba.base_addr + LBA_ERROR_CONFIG);		\
\
    /* Save contents of status control register.  */			\
    status_control = READ_REG32(d->hba.base_addr + LBA_STAT_CTL);		\
\
    /* For LBA rev 2.0, 2.1, 2.2, and 3.0, we must disable DMA		\
    ** arbitration for full bus walks.					\
    */									\
	/* Save contents of arb mask register. */			\
	arb_mask = READ_REG32(d->hba.base_addr + LBA_ARB_MASK);		\
\
	/*								\
	 * Turn off all device arbitration bits (i.e. everything	\
	 * except arbitration enable bit).				\
	 */								\
	WRITE_REG32(0x1, d->hba.base_addr + LBA_ARB_MASK);		\
\
    /*									\
     * Set the smart mode bit so that master aborts don't cause		\
     * LBA to go into PCI fatal mode (required).			\
     */									\
    WRITE_REG32(error_config | LBA_SMART_MODE, d->hba.base_addr + LBA_ERROR_CONFIG);	\
}


#define LBA_CFG_PROBE(d, tok) {				\
    /*									\
     * Setup Vendor ID write and read back the address register		\
     * to make sure that LBA is the bus master.				\
     */									\
    WRITE_REG32(tok | PCI_VENDOR_ID, (d)->hba.base_addr + LBA_PCI_CFG_ADDR);\
    /*									\
     * Read address register to ensure that LBA is the bus master,	\
     * which implies that DMA traffic has stopped when DMA arb is off.	\
     */									\
    lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
    /*									\
     * Generate a cfg write cycle (will have no affect on		\
     * Vendor ID register since read-only).				\
     */									\
    WRITE_REG32(~0, (d)->hba.base_addr + LBA_PCI_CFG_DATA);		\
    /*									\
     * Make sure write has completed before proceeding further,		\
     * i.e. before setting clear enable.				\
     */									\
    lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
}


/*
 * HPREVISIT:
 *   -- Can't tell if config cycle got the error.
 *
 *		OV bit is broken until rev 4.0, so can't use OV bit and
 *		LBA_ERROR_LOG_ADDR to tell if error belongs to config cycle.
 *
 *		As of rev 4.0, no longer need the error check.
 *
 *   -- Even if we could tell, we still want to return -1
 *	for **ANY** error (not just master abort).
 *
 *   -- Only clear non-fatal errors (we don't want to bring
 *	LBA out of pci-fatal mode).
 *
 *		Actually, there is still a race in which
 *		we could be clearing a fatal error.  We will
 *		live with this during our initial bus walk
 *		until rev 4.0 (no driver activity during
 *		initial bus walk).  The initial bus walk
 *		has race conditions concerning the use of
 *		smart mode as well.
 */

#define LBA_MASTER_ABORT_ERROR 0xc
#define LBA_FATAL_ERROR 0x10

#define LBA_CFG_MASTER_ABORT_CHECK(d, base, tok, error) {		\
    u32 error_status = 0;						\
    /*									\
     * Set clear enable (CE) bit. Unset by HW when new			\
     * errors are logged -- LBA HW ERS section 14.3.3).		\
     */									\
    WRITE_REG32(status_control | CLEAR_ERRLOG_ENABLE, base + LBA_STAT_CTL); \
    error_status = READ_REG32(base + LBA_ERROR_STATUS);		\
    if ((error_status & 0x1f) != 0) {					\
	/*								\
	 * Fail the config read request.				\
	 */								\
	error = 1;							\
	if ((error_status & LBA_FATAL_ERROR) == 0) {			\
	    /*								\
	     * Clear error status (if fatal bit not set) by setting	\
	     * clear error log bit (CL).				\
	     */								\
	    WRITE_REG32(status_control | CLEAR_ERRLOG, base + LBA_STAT_CTL); \
	}								\
    }									\
}

#define LBA_CFG_TR4_ADDR_SETUP(d, addr)					\
	WRITE_REG32(((addr) & ~3), (d)->hba.base_addr + LBA_PCI_CFG_ADDR);

#define LBA_CFG_ADDR_SETUP(d, addr) {					\
    WRITE_REG32(((addr) & ~3), (d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
    /*									\
     * Read address register to ensure that LBA is the bus master,	\
     * which implies that DMA traffic has stopped when DMA arb is off.	\
     */									\
    lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
}


#define LBA_CFG_RESTORE(d, base) {					\
    /*									\
     * Restore status control register (turn off clear enable).		\
     */									\
    WRITE_REG32(status_control, base + LBA_STAT_CTL);			\
    /*									\
     * Restore error config register (turn off smart mode).		\
     */									\
    WRITE_REG32(error_config, base + LBA_ERROR_CONFIG);			\
	/*								\
	 * Restore arb mask register (reenables DMA arbitration).	\
	 */								\
	WRITE_REG32(arb_mask, base + LBA_ARB_MASK);			\
}



static unsigned int
lba_rd_cfg(struct lba_device *d, u32 tok, u8 reg, u32 size)
{
	u32 data = ~0U;
	int error = 0;
	u32 arb_mask = 0;	/* used by LBA_CFG_SETUP/RESTORE */
	u32 error_config = 0;	/* used by LBA_CFG_SETUP/RESTORE */
	u32 status_control = 0;	/* used by LBA_CFG_SETUP/RESTORE */

	LBA_CFG_SETUP(d, tok);
	LBA_CFG_PROBE(d, tok);
	LBA_CFG_MASTER_ABORT_CHECK(d, d->hba.base_addr, tok, error);
	if (!error) {
		void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;

		LBA_CFG_ADDR_SETUP(d, tok | reg);
		switch (size) {
		case 1: data = (u32) READ_REG8(data_reg + (reg & 3)); break;
		case 2: data = (u32) READ_REG16(data_reg+ (reg & 2)); break;
		case 4: data = READ_REG32(data_reg); break;
		}
	}
	LBA_CFG_RESTORE(d, d->hba.base_addr);
	return(data);
}


static int elroy_cfg_read(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 *data)
{
	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
	u32 local_bus = (bus->parent == NULL) ? 0 : bus->secondary;
	u32 tok = LBA_CFG_TOK(local_bus, devfn);
	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;

	if ((pos > 255) || (devfn > 255))
		return -EINVAL;

/* FIXME: B2K/C3600 workaround is always use old method... */
	/* if (!LBA_SKIP_PROBE(d)) */ {
		/* original - Generate config cycle on broken elroy
		  with risk we will miss PCI bus errors. */
		*data = lba_rd_cfg(d, tok, pos, size);
		DBG_CFG("%s(%x+%2x) -> 0x%x (a)\n", __FUNCTION__, tok, pos, *data);
		return 0;
	}

	if (LBA_SKIP_PROBE(d) && !lba_device_present(bus->secondary, devfn, d)) {
		DBG_CFG("%s(%x+%2x) -> -1 (b)\n", __FUNCTION__, tok, pos);
		/* either don't want to look or know device isn't present. */
		*data = ~0U;
		return(0);
	}

	/* Basic Algorithm
	** Should only get here on fully working LBA rev.
	** This is how simple the code should have been.
	*/
	LBA_CFG_ADDR_SETUP(d, tok | pos);
	switch(size) {
	case 1: *data = READ_REG8 (data_reg + (pos & 3)); break;
	case 2: *data = READ_REG16(data_reg + (pos & 2)); break;
	case 4: *data = READ_REG32(data_reg); break;
	}
	DBG_CFG("%s(%x+%2x) -> 0x%x (c)\n", __FUNCTION__, tok, pos, *data);
	return 0;
}


static void
lba_wr_cfg(struct lba_device *d, u32 tok, u8 reg, u32 data, u32 size)
{
	int error = 0;
	u32 arb_mask = 0;
	u32 error_config = 0;
	u32 status_control = 0;
	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;

	LBA_CFG_SETUP(d, tok);
	LBA_CFG_ADDR_SETUP(d, tok | reg);
	switch (size) {
	case 1: WRITE_REG8 (data, data_reg + (reg & 3)); break;
	case 2: WRITE_REG16(data, data_reg + (reg & 2)); break;
	case 4: WRITE_REG32(data, data_reg);             break;
	}
	LBA_CFG_MASTER_ABORT_CHECK(d, d->hba.base_addr, tok, error);
	LBA_CFG_RESTORE(d, d->hba.base_addr);
}


/*
 * LBA 4.0 config write code implements non-postable semantics
 * by doing a read of CONFIG ADDR after the write.
 */

static int elroy_cfg_write(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 data)
{
	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
	u32 local_bus = (bus->parent == NULL) ? 0 : bus->secondary;
	u32 tok = LBA_CFG_TOK(local_bus,devfn);

	if ((pos > 255) || (devfn > 255))
		return -EINVAL;

	if (!LBA_SKIP_PROBE(d)) {
		/* Original Workaround */
		lba_wr_cfg(d, tok, pos, (u32) data, size);
		DBG_CFG("%s(%x+%2x) = 0x%x (a)\n", __FUNCTION__, tok, pos,data);
		return 0;
	}

	if (LBA_SKIP_PROBE(d) && (!lba_device_present(bus->secondary, devfn, d))) {
		DBG_CFG("%s(%x+%2x) = 0x%x (b)\n", __FUNCTION__, tok, pos,data);
		return 1; /* New Workaround */
	}

	DBG_CFG("%s(%x+%2x) = 0x%x (c)\n", __FUNCTION__, tok, pos, data);

	/* Basic Algorithm */
	LBA_CFG_ADDR_SETUP(d, tok | pos);
	switch(size) {
	case 1: WRITE_REG8 (data, d->hba.base_addr + LBA_PCI_CFG_DATA + (pos & 3));
		   break;
	case 2: WRITE_REG16(data, d->hba.base_addr + LBA_PCI_CFG_DATA + (pos & 2));
		   break;
	case 4: WRITE_REG32(data, d->hba.base_addr + LBA_PCI_CFG_DATA);
		   break;
	}
	/* flush posted write */
	lba_t32 = READ_REG32(d->hba.base_addr + LBA_PCI_CFG_ADDR);
	return 0;
}


static struct pci_ops elroy_cfg_ops = {
	.read =		elroy_cfg_read,
	.write =	elroy_cfg_write,
};

/*
 * The mercury_cfg_ops are slightly misnamed; they're also used for Elroy
 * TR4.0 as no additional bugs were found in this areea between Elroy and
 * Mercury
 */

static int mercury_cfg_read(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 *data)
{
	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
	u32 local_bus = (bus->parent == NULL) ? 0 : bus->secondary;
	u32 tok = LBA_CFG_TOK(local_bus, devfn);
	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;

	if ((pos > 255) || (devfn > 255))
		return -EINVAL;

	LBA_CFG_TR4_ADDR_SETUP(d, tok | pos);
	switch(size) {
	case 1:
		*data = READ_REG8(data_reg + (pos & 3));
		break;
	case 2:
		*data = READ_REG16(data_reg + (pos & 2));
		break;
	case 4:
		*data = READ_REG32(data_reg);             break;
		break;
	}

	DBG_CFG("mercury_cfg_read(%x+%2x) -> 0x%x\n", tok, pos, *data);
	return 0;
}

/*
 * LBA 4.0 config write code implements non-postable semantics
 * by doing a read of CONFIG ADDR after the write.
 */

static int mercury_cfg_write(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 data)
{
	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
	u32 local_bus = (bus->parent == NULL) ? 0 : bus->secondary;
	u32 tok = LBA_CFG_TOK(local_bus,devfn);

	if ((pos > 255) || (devfn > 255))
		return -EINVAL;

	DBG_CFG("%s(%x+%2x) <- 0x%x (c)\n", __FUNCTION__, tok, pos, data);

	LBA_CFG_TR4_ADDR_SETUP(d, tok | pos);
	switch(size) {
	case 1:
		WRITE_REG8 (data, data_reg + (pos & 3));
		break;
	case 2:
		WRITE_REG16(data, data_reg + (pos & 2));
		break;
	case 4:
		WRITE_REG32(data, data_reg);
		break;
	}

	/* flush posted write */
	lba_t32 = READ_U32(d->hba.base_addr + LBA_PCI_CFG_ADDR);
	return 0;
}

static struct pci_ops mercury_cfg_ops = {
	.read =		mercury_cfg_read,
	.write =	mercury_cfg_write,
};


static void
lba_bios_init(void)
{
	DBG(MODULE_NAME ": lba_bios_init\n");
}


#ifdef CONFIG_64BIT

/*
 * truncate_pat_collision:  Deal with overlaps or outright collisions
 *			between PAT PDC reported ranges.
 *
 *   Broken PA8800 firmware will report lmmio range that
 *   overlaps with CPU HPA. Just truncate the lmmio range.
 *
 *   BEWARE: conflicts with this lmmio range may be an
 *   elmmio range which is pointing down another rope.
 *
 *  FIXME: only deals with one collision per range...theoretically we
 *  could have several. Supporting more than one collision will get messy.
 */
static unsigned long
truncate_pat_collision(struct resource *root, struct resource *new)
{
	unsigned long start = new->start;
	unsigned long end = new->end;
	struct resource *tmp = root->child;

	if (end <= start || start < root->start || !tmp)
		return 0;

	/* find first overlap */
	while (tmp && tmp->end < start)
		tmp = tmp->sibling;

	/* no entries overlap */
	if (!tmp)  return 0;

	/* found one that starts behind the new one
	** Don't need to do anything.
	*/
	if (tmp->start >= end) return 0;

	if (tmp->start <= start) {
		/* "front" of new one overlaps */
		new->start = tmp->end + 1;

		if (tmp->end >= end) {
			/* AACCKK! totally overlaps! drop this range. */
			return 1;
		}
	} 

	if (tmp->end < end ) {
		/* "end" of new one overlaps */
		new->end = tmp->start - 1;
	}

	printk(KERN_WARNING "LBA: Truncating lmmio_space [%lx/%lx] "
					"to [%lx,%lx]\n",
			start, end,
			(long)new->start, (long)new->end );

	return 0;	/* truncation successful */
}

#else
#define truncate_pat_collision(r,n)  (0)
#endif

/*
** The algorithm is generic code.
** But it needs to access local data structures to get the IRQ base.
** Could make this a "pci_fixup_irq(bus, region)" but not sure
** it's worth it.
**
** Called by do_pci_scan_bus() immediately after each PCI bus is walked.
** Resources aren't allocated until recursive buswalk below HBA is completed.
*/
static void
lba_fixup_bus(struct pci_bus *bus)
{
	struct list_head *ln;
#ifdef FBB_SUPPORT
	u16 status;
#endif
	struct lba_device *ldev = LBA_DEV(parisc_walk_tree(bus->bridge));
	int lba_portbase = HBA_PORT_BASE(ldev->hba.hba_num);

	DBG("lba_fixup_bus(0x%p) bus %d platform_data 0x%p\n",
		bus, bus->secondary, bus->bridge->platform_data);

	/*
	** Properly Setup MMIO resources for this bus.
	** pci_alloc_primary_bus() mangles this.
	*/
	if (bus->self) {
		int i;
		/* PCI-PCI Bridge */
		pci_read_bridge_bases(bus);
		for (i = PCI_BRIDGE_RESOURCES; i < PCI_NUM_RESOURCES; i++) {
			pci_claim_resource(bus->self, i);
		}
	} else {
		/* Host-PCI Bridge */
		int err, i;

		DBG("lba_fixup_bus() %s [%lx/%lx]/%lx\n",
			ldev->hba.io_space.name,
			ldev->hba.io_space.start, ldev->hba.io_space.end,
			ldev->hba.io_space.flags);
		DBG("lba_fixup_bus() %s [%lx/%lx]/%lx\n",
			ldev->hba.lmmio_space.name,
			ldev->hba.lmmio_space.start, ldev->hba.lmmio_space.end,
			ldev->hba.lmmio_space.flags);

		err = request_resource(&ioport_resource, &(ldev->hba.io_space));
		if (err < 0) {
			lba_dump_res(&ioport_resource, 2);
			BUG();
		}
		/* advertize Host bridge resources to PCI bus */
		bus->resource[0] = &(ldev->hba.io_space);
		i = 1;

		if (ldev->hba.elmmio_space.start) {
			err = request_resource(&iomem_resource,
					&(ldev->hba.elmmio_space));
			if (err < 0) {

				printk("FAILED: lba_fixup_bus() request for "
						"elmmio_space [%lx/%lx]\n",
						(long)ldev->hba.elmmio_space.start,
						(long)ldev->hba.elmmio_space.end);

				/* lba_dump_res(&iomem_resource, 2); */
				/* BUG(); */
			} else
				bus->resource[i++] = &(ldev->hba.elmmio_space);
		}


		/*   Overlaps with elmmio can (and should) fail here.
		 *   We will prune (or ignore) the distributed range.
		 *
		 *   FIXME: SBA code should register all elmmio ranges first.
		 *      that would take care of elmmio ranges routed
		 *	to a different rope (already discovered) from
		 *	getting registered *after* LBA code has already
		 *	registered it's distributed lmmio range.
		 */
		if (truncate_pat_collision(&iomem_resource,
				       	&(ldev->hba.lmmio_space))) {

			printk(KERN_WARNING "LBA: lmmio_space [%lx/%lx] duplicate!\n",
					(long)ldev->hba.lmmio_space.start,
					(long)ldev->hba.lmmio_space.end);
		} else {
			err = request_resource(&iomem_resource, &(ldev->hba.lmmio_space));
			if (err < 0) {
				printk(KERN_ERR "FAILED: lba_fixup_bus() request for "
					"lmmio_space [%lx/%lx]\n",
					(long)ldev->hba.lmmio_space.start,
					(long)ldev->hba.lmmio_space.end);
			} else
				bus->resource[i++] = &(ldev->hba.lmmio_space);
		}

#ifdef CONFIG_64BIT
		/* GMMIO is  distributed range. Every LBA/Rope gets part it. */
		if (ldev->hba.gmmio_space.flags) {
			err = request_resource(&iomem_resource, &(ldev->hba.gmmio_space));
			if (err < 0) {
				printk("FAILED: lba_fixup_bus() request for "
					"gmmio_space [%lx/%lx]\n",
					(long)ldev->hba.gmmio_space.start,
					(long)ldev->hba.gmmio_space.end);
				lba_dump_res(&iomem_resource, 2);
				BUG();
			}
			bus->resource[i++] = &(ldev->hba.gmmio_space);
		}
#endif

	}

	list_for_each(ln, &bus->devices) {
		int i;
		struct pci_dev *dev = pci_dev_b(ln);

		DBG("lba_fixup_bus() %s\n", pci_name(dev));

		/* Virtualize Device/Bridge Resources. */
		for (i = 0; i < PCI_BRIDGE_RESOURCES; i++) {
			struct resource *res = &dev->resource[i];

			/* If resource not allocated - skip it */
			if (!res->start)
				continue;

			if (res->flags & IORESOURCE_IO) {
				DBG("lba_fixup_bus() I/O Ports [%lx/%lx] -> ",
					res->start, res->end);
				res->start |= lba_portbase;
				res->end   |= lba_portbase;
				DBG("[%lx/%lx]\n", res->start, res->end);
			} else if (res->flags & IORESOURCE_MEM) {
				/*
				** Convert PCI (IO_VIEW) addresses to
				** processor (PA_VIEW) addresses
				 */
				DBG("lba_fixup_bus() MMIO [%lx/%lx] -> ",
					res->start, res->end);
				res->start = PCI_HOST_ADDR(HBA_DATA(ldev), res->start);
				res->end   = PCI_HOST_ADDR(HBA_DATA(ldev), res->end);
				DBG("[%lx/%lx]\n", res->start, res->end);
			} else {
				DBG("lba_fixup_bus() WTF? 0x%lx [%lx/%lx] XXX",
					res->flags, res->start, res->end);
			}

			/*
			** FIXME: this will result in whinging for devices
			** that share expansion ROMs (think quad tulip), but
			** isn't harmful.
			*/
			pci_claim_resource(dev, i);
		}

#ifdef FBB_SUPPORT
		/*
		** If one device does not support FBB transfers,
		** No one on the bus can be allowed to use them.
		*/
		(void) pci_read_config_word(dev, PCI_STATUS, &status);
		bus->bridge_ctl &= ~(status & PCI_STATUS_FAST_BACK);
#endif

                /*
		** P2PB's have no IRQs. ignore them.
		*/
		if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI)
			continue;

		/* Adjust INTERRUPT_LINE for this dev */
		iosapic_fixup_irq(ldev->iosapic_obj, dev);
	}

#ifdef FBB_SUPPORT
/* FIXME/REVISIT - finish figuring out to set FBB on both
** pci_setup_bridge() clobbers PCI_BRIDGE_CONTROL.
** Can't fixup here anyway....garr...
*/
	if (fbb_enable) {
		if (bus->self) {
			u8 control;
			/* enable on PPB */
			(void) pci_read_config_byte(bus->self, PCI_BRIDGE_CONTROL, &control);
			(void) pci_write_config_byte(bus->self, PCI_BRIDGE_CONTROL, control | PCI_STATUS_FAST_BACK);

		} else {
			/* enable on LBA */
		}
		fbb_enable = PCI_COMMAND_FAST_BACK;
	}

	/* Lastly enable FBB/PERR/SERR on all devices too */
	list_for_each(ln, &bus->devices) {
		(void) pci_read_config_word(dev, PCI_COMMAND, &status);
		status |= PCI_COMMAND_PARITY | PCI_COMMAND_SERR | fbb_enable;
		(void) pci_write_config_word(dev, PCI_COMMAND, status);
	}
#endif
}


struct pci_bios_ops lba_bios_ops = {
	.init =		lba_bios_init,
	.fixup_bus =	lba_fixup_bus,
};




/*******************************************************
**
** LBA Sprockets "I/O Port" Space Accessor Functions
**
** This set of accessor functions is intended for use with
** "legacy firmware" (ie Sprockets on Allegro/Forte boxes).
**
** Many PCI devices don't require use of I/O port space (eg Tulip,
** NCR720) since they export the same registers to both MMIO and
** I/O port space. In general I/O port space is slower than
** MMIO since drivers are designed so PIO writes can be posted.
**
********************************************************/

#define LBA_PORT_IN(size, mask) \
static u##size lba_astro_in##size (struct pci_hba_data *d, u16 addr) \
{ \
	u##size t; \
	t = READ_REG##size(astro_iop_base + addr); \
	DBG_PORT(" 0x%x\n", t); \
	return (t); \
}

LBA_PORT_IN( 8, 3)
LBA_PORT_IN(16, 2)
LBA_PORT_IN(32, 0)



/*
** BUG X4107:  Ordering broken - DMA RD return can bypass PIO WR
**
** Fixed in Elroy 2.2. The READ_U32(..., LBA_FUNC_ID) below is
** guarantee non-postable completion semantics - not avoid X4107.
** The READ_U32 only guarantees the write data gets to elroy but
** out to the PCI bus. We can't read stuff from I/O port space
** since we don't know what has side-effects. Attempting to read
** from configuration space would be suicidal given the number of
** bugs in that elroy functionality.
**
**      Description:
**          DMA read results can improperly pass PIO writes (X4107).  The
**          result of this bug is that if a processor modifies a location in
**          memory after having issued PIO writes, the PIO writes are not
**          guaranteed to be completed before a PCI device is allowed to see
**          the modified data in a DMA read.
**
**          Note that IKE bug X3719 in TR1 IKEs will result in the same
**          symptom.
**
**      Workaround:
**          The workaround for this bug is to always follow a PIO write with
**          a PIO read to the same bus before starting DMA on that PCI bus.
**
*/
#define LBA_PORT_OUT(size, mask) \
static void lba_astro_out##size (struct pci_hba_data *d, u16 addr, u##size val) \
{ \
	DBG_PORT("%s(0x%p, 0x%x, 0x%x)\n", __FUNCTION__, d, addr, val); \
	WRITE_REG##size(val, astro_iop_base + addr); \
	if (LBA_DEV(d)->hw_rev < 3) \
		lba_t32 = READ_U32(d->base_addr + LBA_FUNC_ID); \
}

LBA_PORT_OUT( 8, 3)
LBA_PORT_OUT(16, 2)
LBA_PORT_OUT(32, 0)


static struct pci_port_ops lba_astro_port_ops = {
	.inb =	lba_astro_in8,
	.inw =	lba_astro_in16,
	.inl =	lba_astro_in32,
	.outb =	lba_astro_out8,
	.outw =	lba_astro_out16,
	.outl =	lba_astro_out32
};


#ifdef CONFIG_64BIT
#define PIOP_TO_GMMIO(lba, addr) \
	((lba)->iop_base + (((addr)&0xFFFC)<<10) + ((addr)&3))

/*******************************************************
**
** LBA PAT "I/O Port" Space Accessor Functions
**
** This set of accessor functions is intended for use with
** "PAT PDC" firmware (ie Prelude/Rhapsody/Piranha boxes).
**
** This uses the PIOP space located in the first 64MB of GMMIO.
** Each rope gets a full 64*KB* (ie 4 bytes per page) this way.
** bits 1:0 stay the same.  bits 15:2 become 25:12.
** Then add the base and we can generate an I/O Port cycle.
********************************************************/
#undef LBA_PORT_IN
#define LBA_PORT_IN(size, mask) \
static u##size lba_pat_in##size (struct pci_hba_data *l, u16 addr) \
{ \
	u##size t; \
	DBG_PORT("%s(0x%p, 0x%x) ->", __FUNCTION__, l, addr); \
	t = READ_REG##size(PIOP_TO_GMMIO(LBA_DEV(l), addr)); \
	DBG_PORT(" 0x%x\n", t); \
	return (t); \
}

LBA_PORT_IN( 8, 3)
LBA_PORT_IN(16, 2)
LBA_PORT_IN(32, 0)


#undef LBA_PORT_OUT
#define LBA_PORT_OUT(size, mask) \
static void lba_pat_out##size (struct pci_hba_data *l, u16 addr, u##size val) \
{ \
	void __iomem *where = PIOP_TO_GMMIO(LBA_DEV(l), addr); \
	DBG_PORT("%s(0x%p, 0x%x, 0x%x)\n", __FUNCTION__, l, addr, val); \
	WRITE_REG##size(val, where); \
	/* flush the I/O down to the elroy at least */ \
	lba_t32 = READ_U32(l->base_addr + LBA_FUNC_ID); \
}

LBA_PORT_OUT( 8, 3)
LBA_PORT_OUT(16, 2)
LBA_PORT_OUT(32, 0)


static struct pci_port_ops lba_pat_port_ops = {
	.inb =	lba_pat_in8,
	.inw =	lba_pat_in16,
	.inl =	lba_pat_in32,
	.outb =	lba_pat_out8,
	.outw =	lba_pat_out16,
	.outl =	lba_pat_out32
};



/*
** make range information from PDC available to PCI subsystem.
** We make the PDC call here in order to get the PCI bus range
** numbers. The rest will get forwarded in pcibios_fixup_bus().
** We don't have a struct pci_bus assigned to us yet.
*/
static void
lba_pat_resources(struct parisc_device *pa_dev, struct lba_device *lba_dev)
{
	unsigned long bytecnt;
	pdc_pat_cell_mod_maddr_block_t pa_pdc_cell;	/* PA_VIEW */
	pdc_pat_cell_mod_maddr_block_t io_pdc_cell;	/* IO_VIEW */
	long io_count;
	long status;	/* PDC return status */
	long pa_count;
	int i;

	/* return cell module (IO view) */
	status = pdc_pat_cell_module(&bytecnt, pa_dev->pcell_loc, pa_dev->mod_index,
				PA_VIEW, & pa_pdc_cell);
	pa_count = pa_pdc_cell.mod[1];

	status |= pdc_pat_cell_module(&bytecnt, pa_dev->pcell_loc, pa_dev->mod_index,
				IO_VIEW, &io_pdc_cell);
	io_count = io_pdc_cell.mod[1];

	/* We've already done this once for device discovery...*/
	if (status != PDC_OK) {