lba_pci.c

					io->start;
				r = &lba_dev->hba.lmmio_space;
				r->name = lba_dev->hba.lmmio_name;
			} else if (!lba_dev->hba.elmmio_space.start) {
				sprintf(lba_dev->hba.elmmio_name,
						"PCI%02x ELMMIO",
						(int)lba_dev->hba.bus_num.start);
				r = &lba_dev->hba.elmmio_space;
				r->name = lba_dev->hba.elmmio_name;
			} else {
				printk(KERN_WARNING MODULE_NAME
					" only supports 2 LMMIO resources!\n");
				break;
			}

			r->start  = p->start;
			r->end    = p->end;
			r->flags  = IORESOURCE_MEM;
			r->parent = r->sibling = r->child = NULL;
			break;

		case PAT_GMMIO:
			/* MMIO space > 4GB phys addr; for 64-bit BAR */
			sprintf(lba_dev->hba.gmmio_name, "PCI%02x GMMIO",
					(int)lba_dev->hba.bus_num.start);
			r = &lba_dev->hba.gmmio_space;
			r->name  = lba_dev->hba.gmmio_name;
			r->start  = p->start;
			r->end    = p->end;
			r->flags  = IORESOURCE_MEM;
			r->parent = r->sibling = r->child = NULL;
			break;

		case PAT_NPIOP:
			printk(KERN_WARNING MODULE_NAME
				" range[%d] : ignoring NPIOP (0x%lx)\n",
				i, p->start);
			break;

		case PAT_PIOP:
			/*
			** Postable I/O port space is per PCI host adapter.
			** base of 64MB PIOP region
			*/
			lba_dev->iop_base = ioremap_nocache(p->start, 64 * 1024 * 1024);

			sprintf(lba_dev->hba.io_name, "PCI%02x Ports",
					(int)lba_dev->hba.bus_num.start);
			r = &lba_dev->hba.io_space;
			r->name  = lba_dev->hba.io_name;
			r->start  = HBA_PORT_BASE(lba_dev->hba.hba_num);
			r->end    = r->start + HBA_PORT_SPACE_SIZE - 1;
			r->flags  = IORESOURCE_IO;
			r->parent = r->sibling = r->child = NULL;
			break;

		default:
			printk(KERN_WARNING MODULE_NAME
				" range[%d] : unknown pat range type (0x%lx)\n",
				i, p->type & 0xff);
			break;
		}
	}

	kfree(pa_pdc_cell);
	kfree(io_pdc_cell);
}
#else
/* keep compiler from complaining about missing declarations */
#define lba_pat_port_ops lba_astro_port_ops
#define lba_pat_resources(pa_dev, lba_dev)
#endif	/* CONFIG_64BIT */


extern void sba_distributed_lmmio(struct parisc_device *, struct resource *);
extern void sba_directed_lmmio(struct parisc_device *, struct resource *);


static void
lba_legacy_resources(struct parisc_device *pa_dev, struct lba_device *lba_dev)
{
	struct resource *r;
	int lba_num;

	lba_dev->hba.lmmio_space_offset = PCI_F_EXTEND;

	/*
	** With "legacy" firmware, the lowest byte of FW_SCRATCH
	** represents bus->secondary and the second byte represents
	** bus->subsidiary (i.e. highest PPB programmed by firmware).
	** PCI bus walk *should* end up with the same result.
	** FIXME: But we don't have sanity checks in PCI or LBA.
	*/
	lba_num = READ_REG32(lba_dev->hba.base_addr + LBA_FW_SCRATCH);
	r = &(lba_dev->hba.bus_num);
	r->name = "LBA PCI Busses";
	r->start = lba_num & 0xff;
	r->end = (lba_num>>8) & 0xff;

	/* Set up local PCI Bus resources - we don't need them for
	** Legacy boxes but it's nice to see in /proc/iomem.
	*/
	r = &(lba_dev->hba.lmmio_space);
	sprintf(lba_dev->hba.lmmio_name, "PCI%02x LMMIO",
					(int)lba_dev->hba.bus_num.start);
	r->name  = lba_dev->hba.lmmio_name;

#if 1
	/* We want the CPU -> IO routing of addresses.
	 * The SBA BASE/MASK registers control CPU -> IO routing.
	 * Ask SBA what is routed to this rope/LBA.
	 */
	sba_distributed_lmmio(pa_dev, r);
#else
	/*
	 * The LBA BASE/MASK registers control IO -> System routing.
	 *
	 * The following code works but doesn't get us what we want.
	 * Well, only because firmware (v5.0) on C3000 doesn't program
	 * the LBA BASE/MASE registers to be the exact inverse of 
	 * the corresponding SBA registers. Other Astro/Pluto
	 * based platform firmware may do it right.
	 *
	 * Should someone want to mess with MSI, they may need to
	 * reprogram LBA BASE/MASK registers. Thus preserve the code
	 * below until MSI is known to work on C3000/A500/N4000/RP3440.
	 *
	 * Using the code below, /proc/iomem shows:
	 * ...
	 * f0000000-f0ffffff : PCI00 LMMIO
	 *   f05d0000-f05d0000 : lcd_data
	 *   f05d0008-f05d0008 : lcd_cmd
	 * f1000000-f1ffffff : PCI01 LMMIO
	 * f4000000-f4ffffff : PCI02 LMMIO
	 *   f4000000-f4001fff : sym53c8xx
	 *   f4002000-f4003fff : sym53c8xx
	 *   f4004000-f40043ff : sym53c8xx
	 *   f4005000-f40053ff : sym53c8xx
	 *   f4007000-f4007fff : ohci_hcd
	 *   f4008000-f40083ff : tulip
	 * f6000000-f6ffffff : PCI03 LMMIO
	 * f8000000-fbffffff : PCI00 ELMMIO
	 *   fa100000-fa4fffff : stifb mmio
	 *   fb000000-fb1fffff : stifb fb
	 *
	 * But everything listed under PCI02 actually lives under PCI00.
	 * This is clearly wrong.
	 *
	 * Asking SBA how things are routed tells the correct story:
	 * LMMIO_BASE/MASK/ROUTE f4000001 fc000000 00000000
	 * DIR0_BASE/MASK/ROUTE fa000001 fe000000 00000006
	 * DIR1_BASE/MASK/ROUTE f9000001 ff000000 00000004
	 * DIR2_BASE/MASK/ROUTE f0000000 fc000000 00000000
	 * DIR3_BASE/MASK/ROUTE f0000000 fc000000 00000000
	 *
	 * Which looks like this in /proc/iomem:
	 * f4000000-f47fffff : PCI00 LMMIO
	 *   f4000000-f4001fff : sym53c8xx
	 *   ...[deteled core devices - same as above]...
	 *   f4008000-f40083ff : tulip
	 * f4800000-f4ffffff : PCI01 LMMIO
	 * f6000000-f67fffff : PCI02 LMMIO
	 * f7000000-f77fffff : PCI03 LMMIO
	 * f9000000-f9ffffff : PCI02 ELMMIO
	 * fa000000-fbffffff : PCI03 ELMMIO
	 *   fa100000-fa4fffff : stifb mmio
	 *   fb000000-fb1fffff : stifb fb
	 *
	 * ie all Built-in core are under now correctly under PCI00.
	 * The "PCI02 ELMMIO" directed range is for:
	 *  +-[02]---03.0  3Dfx Interactive, Inc. Voodoo 2
	 *
	 * All is well now.
	 */
	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_LMMIO_BASE);
	if (r->start & 1) {
		unsigned long rsize;

		r->flags = IORESOURCE_MEM;
		/* mmio_mask also clears Enable bit */
		r->start &= mmio_mask;
		r->start = PCI_HOST_ADDR(HBA_DATA(lba_dev), r->start);
		rsize = ~ READ_REG32(lba_dev->hba.base_addr + LBA_LMMIO_MASK);

		/*
		** Each rope only gets part of the distributed range.
		** Adjust "window" for this rope.
		*/
		rsize /= ROPES_PER_IOC;
		r->start += (rsize + 1) * LBA_NUM(pa_dev->hpa.start);
		r->end = r->start + rsize;
	} else {
		r->end = r->start = 0;	/* Not enabled. */
	}
#endif

	/*
	** "Directed" ranges are used when the "distributed range" isn't
	** sufficient for all devices below a given LBA.  Typically devices
	** like graphics cards or X25 may need a directed range when the
	** bus has multiple slots (ie multiple devices) or the device
	** needs more than the typical 4 or 8MB a distributed range offers.
	**
	** The main reason for ignoring it now frigging complications.
	** Directed ranges may overlap (and have precedence) over
	** distributed ranges. Or a distributed range assigned to a unused
	** rope may be used by a directed range on a different rope.
	** Support for graphics devices may require fixing this
	** since they may be assigned a directed range which overlaps
	** an existing (but unused portion of) distributed range.
	*/
	r = &(lba_dev->hba.elmmio_space);
	sprintf(lba_dev->hba.elmmio_name, "PCI%02x ELMMIO",
					(int)lba_dev->hba.bus_num.start);
	r->name  = lba_dev->hba.elmmio_name;

#if 1
	/* See comment which precedes call to sba_directed_lmmio() */
	sba_directed_lmmio(pa_dev, r);
#else
	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_ELMMIO_BASE);

	if (r->start & 1) {
		unsigned long rsize;
		r->flags = IORESOURCE_MEM;
		/* mmio_mask also clears Enable bit */
		r->start &= mmio_mask;
		r->start = PCI_HOST_ADDR(HBA_DATA(lba_dev), r->start);
		rsize = READ_REG32(lba_dev->hba.base_addr + LBA_ELMMIO_MASK);
		r->end = r->start + ~rsize;
	}
#endif

	r = &(lba_dev->hba.io_space);
	sprintf(lba_dev->hba.io_name, "PCI%02x Ports",
					(int)lba_dev->hba.bus_num.start);
	r->name  = lba_dev->hba.io_name;
	r->flags = IORESOURCE_IO;
	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_IOS_BASE) & ~1L;
	r->end   = r->start + (READ_REG32(lba_dev->hba.base_addr + LBA_IOS_MASK) ^ (HBA_PORT_SPACE_SIZE - 1));

	/* Virtualize the I/O Port space ranges */
	lba_num = HBA_PORT_BASE(lba_dev->hba.hba_num);
	r->start |= lba_num;
	r->end   |= lba_num;
}


/**************************************************************************
**
**   LBA initialization code (HW and SW)
**
**   o identify LBA chip itself
**   o initialize LBA chip modes (HardFail)
**   o FIXME: initialize DMA hints for reasonable defaults
**   o enable configuration functions
**   o call pci_register_ops() to discover devs (fixup/fixup_bus get invoked)
**
**************************************************************************/

static int __init
lba_hw_init(struct lba_device *d)
{
	u32 stat;
	u32 bus_reset;	/* PDC_PAT_BUG */

#if 0
	printk(KERN_DEBUG "LBA %lx  STAT_CTL %Lx  ERROR_CFG %Lx  STATUS %Lx DMA_CTL %Lx\n",
		d->hba.base_addr,
		READ_REG64(d->hba.base_addr + LBA_STAT_CTL),
		READ_REG64(d->hba.base_addr + LBA_ERROR_CONFIG),
		READ_REG64(d->hba.base_addr + LBA_ERROR_STATUS),
		READ_REG64(d->hba.base_addr + LBA_DMA_CTL) );
	printk(KERN_DEBUG "	ARB mask %Lx  pri %Lx  mode %Lx  mtlt %Lx\n",
		READ_REG64(d->hba.base_addr + LBA_ARB_MASK),
		READ_REG64(d->hba.base_addr + LBA_ARB_PRI),
		READ_REG64(d->hba.base_addr + LBA_ARB_MODE),
		READ_REG64(d->hba.base_addr + LBA_ARB_MTLT) );
	printk(KERN_DEBUG "	HINT cfg 0x%Lx\n",
		READ_REG64(d->hba.base_addr + LBA_HINT_CFG));
	printk(KERN_DEBUG "	HINT reg ");
	{ int i;
	for (i=LBA_HINT_BASE; i< (14*8 + LBA_HINT_BASE); i+=8)
		printk(" %Lx", READ_REG64(d->hba.base_addr + i));
	}
	printk("\n");
#endif	/* DEBUG_LBA_PAT */

#ifdef CONFIG_64BIT
/*
 * FIXME add support for PDC_PAT_IO "Get slot status" - OLAR support
 * Only N-Class and up can really make use of Get slot status.
 * maybe L-class too but I've never played with it there.
 */
#endif

	/* PDC_PAT_BUG: exhibited in rev 40.48  on L2000 */
	bus_reset = READ_REG32(d->hba.base_addr + LBA_STAT_CTL + 4) & 1;
	if (bus_reset) {
		printk(KERN_DEBUG "NOTICE: PCI bus reset still asserted! (clearing)\n");
	}

	stat = READ_REG32(d->hba.base_addr + LBA_ERROR_CONFIG);
	if (stat & LBA_SMART_MODE) {
		printk(KERN_DEBUG "NOTICE: LBA in SMART mode! (cleared)\n");
		stat &= ~LBA_SMART_MODE;
		WRITE_REG32(stat, d->hba.base_addr + LBA_ERROR_CONFIG);
	}

	/* Set HF mode as the default (vs. -1 mode). */
        stat = READ_REG32(d->hba.base_addr + LBA_STAT_CTL);
	WRITE_REG32(stat | HF_ENABLE, d->hba.base_addr + LBA_STAT_CTL);

	/*
	** Writing a zero to STAT_CTL.rf (bit 0) will clear reset signal
	** if it's not already set. If we just cleared the PCI Bus Reset
	** signal, wait a bit for the PCI devices to recover and setup.
	*/
	if (bus_reset)
		mdelay(pci_post_reset_delay);

	if (0 == READ_REG32(d->hba.base_addr + LBA_ARB_MASK)) {
		/*
		** PDC_PAT_BUG: PDC rev 40.48 on L2000.
		** B2000/C3600/J6000 also have this problem?
		** 
		** Elroys with hot pluggable slots don't get configured
		** correctly if the slot is empty.  ARB_MASK is set to 0
		** and we can't master transactions on the bus if it's
		** not at least one. 0x3 enables elroy and first slot.
		*/
		printk(KERN_DEBUG "NOTICE: Enabling PCI Arbitration\n");
		WRITE_REG32(0x3, d->hba.base_addr + LBA_ARB_MASK);
	}

	/*
	** FIXME: Hint registers are programmed with default hint
	** values by firmware. Hints should be sane even if we
	** can't reprogram them the way drivers want.
	*/
	return 0;
}

/*
 * Unfortunately, when firmware numbers busses, it doesn't take into account
 * Cardbus bridges.  So we have to renumber the busses to suit ourselves.
 * Elroy/Mercury don't actually know what bus number they're attached to;
 * we use bus 0 to indicate the directly attached bus and any other bus
 * number will be taken care of by the PCI-PCI bridge.
 */
static unsigned int lba_next_bus = 0;

/*
 * Determine if lba should claim this chip (return 0) or not (return 1).
 * If so, initialize the chip and tell other partners in crime they
 * have work to do.
 */
static int __init
lba_driver_probe(struct parisc_device *dev)
{
	struct lba_device *lba_dev;
	LIST_HEAD(resources);
	struct pci_bus *lba_bus;
	struct pci_ops *cfg_ops;
	u32 func_class;
	void *tmp_obj;
	char *version;
	void __iomem *addr = ioremap_nocache(dev->hpa.start, 4096);

	/* Read HW Rev First */
	func_class = READ_REG32(addr + LBA_FCLASS);

	if (IS_ELROY(dev)) {	
		func_class &= 0xf;
		switch (func_class) {
		case 0:	version = "TR1.0"; break;
		case 1:	version = "TR2.0"; break;
		case 2:	version = "TR2.1"; break;
		case 3:	version = "TR2.2"; break;
		case 4:	version = "TR3.0"; break;
		case 5:	version = "TR4.0"; break;
		default: version = "TR4+";
		}

		printk(KERN_INFO "Elroy version %s (0x%x) found at 0x%lx\n",
		       version, func_class & 0xf, (long)dev->hpa.start);

		if (func_class < 2) {
			printk(KERN_WARNING "Can't support LBA older than "
				"TR2.1 - continuing under adversity.\n");
		}

#if 0
/* Elroy TR4.0 should work with simple algorithm.
   But it doesn't.  Still missing something. *sigh*
*/
		if (func_class > 4) {
			cfg_ops = &mercury_cfg_ops;
		} else
#endif
		{
			cfg_ops = &elroy_cfg_ops;
		}

	} else if (IS_MERCURY(dev) || IS_QUICKSILVER(dev)) {
		int major, minor;

		func_class &= 0xff;
		major = func_class >> 4, minor = func_class & 0xf;

		/* We could use one printk for both Elroy and Mercury,
                 * but for the mask for func_class.
                 */ 
		printk(KERN_INFO "%s version TR%d.%d (0x%x) found at 0x%lx\n",
		       IS_MERCURY(dev) ? "Mercury" : "Quicksilver", major,
		       minor, func_class, (long)dev->hpa.start);

		cfg_ops = &mercury_cfg_ops;
	} else {
		printk(KERN_ERR "Unknown LBA found at 0x%lx\n",
			(long)dev->hpa.start);
		return -ENODEV;
	}

	/* Tell I/O SAPIC driver we have a IRQ handler/region. */
	tmp_obj = iosapic_register(dev->hpa.start + LBA_IOSAPIC_BASE);

	/* NOTE: PCI devices (e.g. 103c:1005 graphics card) which don't
	**	have an IRT entry will get NULL back from iosapic code.
	*/
	
	lba_dev = kzalloc(sizeof(struct lba_device), GFP_KERNEL);
	if (!lba_dev) {
		printk(KERN_ERR "lba_init_chip - couldn't alloc lba_device\n");
		return(1);
	}


	/* ---------- First : initialize data we already have --------- */

	lba_dev->hw_rev = func_class;
	lba_dev->hba.base_addr = addr;
	lba_dev->hba.dev = dev;
	lba_dev->iosapic_obj = tmp_obj;  /* save interrupt handle */
	lba_dev->hba.iommu = sba_get_iommu(dev);  /* get iommu data */
	parisc_set_drvdata(dev, lba_dev);

	/* ------------ Second : initialize common stuff ---------- */
	pci_bios = &lba_bios_ops;
	pcibios_register_hba(HBA_DATA(lba_dev));
	spin_lock_init(&lba_dev->lba_lock);

	if (lba_hw_init(lba_dev))
		return(1);

	/* ---------- Third : setup I/O Port and MMIO resources  --------- */

	if (is_pdc_pat()) {
		/* PDC PAT firmware uses PIOP region of GMMIO space. */
		pci_port = &lba_pat_port_ops;
		/* Go ask PDC PAT what resources this LBA has */
		lba_pat_resources(dev, lba_dev);
	} else {
		if (!astro_iop_base) {
			/* Sprockets PDC uses NPIOP region */
			astro_iop_base = ioremap_nocache(LBA_PORT_BASE, 64 * 1024);
			pci_port = &lba_astro_port_ops;
		}

		/* Poke the chip a bit for /proc output */
		lba_legacy_resources(dev, lba_dev);
	}

	if (lba_dev->hba.bus_num.start < lba_next_bus)
		lba_dev->hba.bus_num.start = lba_next_bus;

	/*   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,
				   &(lba_dev->hba.lmmio_space))) {
		printk(KERN_WARNING "LBA: lmmio_space [%lx/%lx] duplicate!\n",
				(long)lba_dev->hba.lmmio_space.start,
				(long)lba_dev->hba.lmmio_space.end);
		lba_dev->hba.lmmio_space.flags = 0;
	}

	pci_add_resource_offset(&resources, &lba_dev->hba.io_space,
				HBA_PORT_BASE(lba_dev->hba.hba_num));
	if (lba_dev->hba.elmmio_space.start)
		pci_add_resource_offset(&resources, &lba_dev->hba.elmmio_space,
					lba_dev->hba.lmmio_space_offset);
	if (lba_dev->hba.lmmio_space.flags)
		pci_add_resource_offset(&resources, &lba_dev->hba.lmmio_space,
					lba_dev->hba.lmmio_space_offset);
	if (lba_dev->hba.gmmio_space.flags)
		pci_add_resource(&resources, &lba_dev->hba.gmmio_space);

	dev->dev.platform_data = lba_dev;
	lba_bus = lba_dev->hba.hba_bus =
		pci_create_root_bus(&dev->dev, lba_dev->hba.bus_num.start,
				    cfg_ops, NULL, &resources);
	if (!lba_bus) {
		pci_free_resource_list(&resources);
		return 0;
	}

	lba_bus->busn_res.end = pci_scan_child_bus(lba_bus);

	/* This is in lieu of calling pci_assign_unassigned_resources() */
	if (is_pdc_pat()) {
		/* assign resources to un-initialized devices */

		DBG_PAT("LBA pci_bus_size_bridges()\n");
		pci_bus_size_bridges(lba_bus);

		DBG_PAT("LBA pci_bus_assign_resources()\n");
		pci_bus_assign_resources(lba_bus);

#ifdef DEBUG_LBA_PAT
		DBG_PAT("\nLBA PIOP resource tree\n");
		lba_dump_res(&lba_dev->hba.io_space, 2);
		DBG_PAT("\nLBA LMMIO resource tree\n");
		lba_dump_res(&lba_dev->hba.lmmio_space, 2);
#endif
	}
	pci_enable_bridges(lba_bus);

	/*
	** Once PCI register ops has walked the bus, access to config
	** space is restricted. Avoids master aborts on config cycles.
	** Early LBA revs go fatal on *any* master abort.
	*/
	if (cfg_ops == &elroy_cfg_ops) {
		lba_dev->flags |= LBA_FLAG_SKIP_PROBE;
	}

	lba_next_bus = lba_res->busn_res.end + 1;
	pci_bus_add_devices(lba_bus);

	/* Whew! Finally done! Tell services we got this one covered. */
	return 0;
}

static struct parisc_device_id lba_tbl[] = {
	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, ELROY_HVERS, 0xa },
	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, MERCURY_HVERS, 0xa },
	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, QUICKSILVER_HVERS, 0xa },
	{ 0, }
};

static struct parisc_driver lba_driver = {
	.name =		MODULE_NAME,
	.id_table =	lba_tbl,
	.probe =	lba_driver_probe,
};

/*
** One time initialization to let the world know the LBA was found.
** Must be called exactly once before pci_init().
*/
void __init lba_init(void)
{
	register_parisc_driver(&lba_driver);
}

/*
** Initialize the IBASE/IMASK registers for LBA (Elroy).
** Only called from sba_iommu.c in order to route ranges (MMIO vs DMA).
** sba_iommu is responsible for locking (none needed at init time).
*/
void lba_set_iregs(struct parisc_device *lba, u32 ibase, u32 imask)
{
	void __iomem * base_addr = ioremap_nocache(lba->hpa.start, 4096);

	imask <<= 2;	/* adjust for hints - 2 more bits */

	/* Make sure we aren't trying to set bits that aren't writeable. */
	WARN_ON((ibase & 0x001fffff) != 0);
	WARN_ON((imask & 0x001fffff) != 0);
	
	DBG("%s() ibase 0x%x imask 0x%x\n", __func__, ibase, imask);
	WRITE_REG32( imask, base_addr + LBA_IMASK);
	WRITE_REG32( ibase, base_addr + LBA_IBASE);
	iounmap(base_addr);
}