Linux-2.6.12-rc2v2.6.12-rc2

Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!

1 files changed, 2165 insertions, 0 deletions

diff --git a/drivers/parisc/sba_iommu.c b/drivers/parisc/sba_iommu.c
new file mode 100644
index 00000000000..82ea68b55df
--- /dev/null
+++ b/drivers/parisc/sba_iommu.c
@@ -0,0 +1,2165 @@
+/*
+**  System Bus Adapter (SBA) I/O MMU manager
+**
+**	(c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
+**	(c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
+**	(c) Copyright 2000-2004 Hewlett-Packard Company
+**
+**	Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
+**
+**	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 initializes the IOC (I/O Controller) found on B1000/C3000/
+** J5000/J7000/N-class/L-class machines and their successors.
+**
+** FIXME: add DMA hint support programming in both sba and lba modules.
+*/
+
+#include <linux/config.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/spinlock.h>
+#include <linux/slab.h>
+#include <linux/init.h>
+
+#include <linux/mm.h>
+#include <linux/string.h>
+#include <linux/pci.h>
+
+#include <asm/byteorder.h>
+#include <asm/io.h>
+#include <asm/dma.h>		/* for DMA_CHUNK_SIZE */
+
+#include <asm/hardware.h>	/* for register_parisc_driver() stuff */
+
+#include <linux/proc_fs.h>
+#include <asm/runway.h>		/* for proc_runway_root */
+#include <asm/pdc.h>		/* for PDC_MODEL_* */
+#include <asm/pdcpat.h>		/* for is_pdc_pat() */
+#include <asm/parisc-device.h>
+
+
+/* declared in arch/parisc/kernel/setup.c */
+extern struct proc_dir_entry * proc_mckinley_root;
+
+#define MODULE_NAME "SBA"
+
+#ifdef CONFIG_PROC_FS
+/* depends on proc fs support. But costs CPU performance */
+#undef SBA_COLLECT_STATS
+#endif
+
+/*
+** The number of debug flags is a clue - this code is fragile.
+** Don't even think about messing with it unless you have
+** plenty of 710's to sacrifice to the computer gods. :^)
+*/
+#undef DEBUG_SBA_INIT
+#undef DEBUG_SBA_RUN
+#undef DEBUG_SBA_RUN_SG
+#undef DEBUG_SBA_RESOURCE
+#undef ASSERT_PDIR_SANITY
+#undef DEBUG_LARGE_SG_ENTRIES
+#undef DEBUG_DMB_TRAP
+
+#ifdef DEBUG_SBA_INIT
+#define DBG_INIT(x...)	printk(x)
+#else
+#define DBG_INIT(x...)
+#endif
+
+#ifdef DEBUG_SBA_RUN
+#define DBG_RUN(x...)	printk(x)
+#else
+#define DBG_RUN(x...)
+#endif
+
+#ifdef DEBUG_SBA_RUN_SG
+#define DBG_RUN_SG(x...)	printk(x)
+#else
+#define DBG_RUN_SG(x...)
+#endif
+
+
+#ifdef DEBUG_SBA_RESOURCE
+#define DBG_RES(x...)	printk(x)
+#else
+#define DBG_RES(x...)
+#endif
+
+#if defined(__LP64__) && !defined(CONFIG_PDC_NARROW)
+/* "low end" PA8800 machines use ZX1 chipset */
+#define ZX1_SUPPORT
+#endif
+
+#define SBA_INLINE	__inline__
+
+
+/*
+** The number of pdir entries to "free" before issueing
+** a read to PCOM register to flush out PCOM writes.
+** Interacts with allocation granularity (ie 4 or 8 entries
+** allocated and free'd/purged at a time might make this
+** less interesting).
+*/
+#define DELAYED_RESOURCE_CNT	16
+
+#define DEFAULT_DMA_HINT_REG	0
+
+#define ASTRO_RUNWAY_PORT	0x582
+#define IKE_MERCED_PORT		0x803
+#define REO_MERCED_PORT		0x804
+#define REOG_MERCED_PORT	0x805
+#define PLUTO_MCKINLEY_PORT	0x880
+
+#define SBA_FUNC_ID	0x0000	/* function id */
+#define SBA_FCLASS	0x0008	/* function class, bist, header, rev... */
+
+#define IS_ASTRO(id)		((id)->hversion == ASTRO_RUNWAY_PORT)
+#define IS_IKE(id)		((id)->hversion == IKE_MERCED_PORT)
+#define IS_PLUTO(id)		((id)->hversion == PLUTO_MCKINLEY_PORT)
+
+#define SBA_FUNC_SIZE 4096   /* SBA configuration function reg set */
+
+#define ASTRO_IOC_OFFSET	(32 * SBA_FUNC_SIZE)
+#define PLUTO_IOC_OFFSET	(1 * SBA_FUNC_SIZE)
+/* Ike's IOC's occupy functions 2 and 3 */
+#define IKE_IOC_OFFSET(p)	((p+2) * SBA_FUNC_SIZE)
+
+#define IOC_CTRL          0x8	/* IOC_CTRL offset */
+#define IOC_CTRL_TC       (1 << 0) /* TOC Enable */
+#define IOC_CTRL_CE       (1 << 1) /* Coalesce Enable */
+#define IOC_CTRL_DE       (1 << 2) /* Dillon Enable */
+#define IOC_CTRL_RM       (1 << 8) /* Real Mode */
+#define IOC_CTRL_NC       (1 << 9) /* Non Coherent Mode */
+#define IOC_CTRL_D4       (1 << 11) /* Disable 4-byte coalescing */
+#define IOC_CTRL_DD       (1 << 13) /* Disable distr. LMMIO range coalescing */
+
+#define MAX_IOC		2	/* per Ike. Pluto/Astro only have 1. */
+
+#define ROPES_PER_IOC	8	/* per Ike half or Pluto/Astro */
+
+
+/*
+** Offsets into MBIB (Function 0 on Ike and hopefully Astro)
+** Firmware programs this stuff. Don't touch it.
+*/
+#define LMMIO_DIRECT0_BASE  0x300
+#define LMMIO_DIRECT0_MASK  0x308
+#define LMMIO_DIRECT0_ROUTE 0x310
+
+#define LMMIO_DIST_BASE  0x360
+#define LMMIO_DIST_MASK  0x368
+#define LMMIO_DIST_ROUTE 0x370
+
+#define IOS_DIST_BASE	0x390
+#define IOS_DIST_MASK	0x398
+#define IOS_DIST_ROUTE	0x3A0
+
+#define IOS_DIRECT_BASE	0x3C0
+#define IOS_DIRECT_MASK	0x3C8
+#define IOS_DIRECT_ROUTE 0x3D0
+
+/*
+** Offsets into I/O TLB (Function 2 and 3 on Ike)
+*/
+#define ROPE0_CTL	0x200  /* "regbus pci0" */
+#define ROPE1_CTL	0x208
+#define ROPE2_CTL	0x210
+#define ROPE3_CTL	0x218
+#define ROPE4_CTL	0x220
+#define ROPE5_CTL	0x228
+#define ROPE6_CTL	0x230
+#define ROPE7_CTL	0x238
+
+#define HF_ENABLE	0x40
+
+
+#define IOC_IBASE	0x300	/* IO TLB */
+#define IOC_IMASK	0x308
+#define IOC_PCOM	0x310
+#define IOC_TCNFG	0x318
+#define IOC_PDIR_BASE	0x320
+
+/* AGP GART driver looks for this */
+#define SBA_IOMMU_COOKIE    0x0000badbadc0ffeeUL
+
+
+/*
+** IOC supports 4/8/16/64KB page sizes (see TCNFG register)
+** It's safer (avoid memory corruption) to keep DMA page mappings
+** equivalently sized to VM PAGE_SIZE.
+**
+** We really can't avoid generating a new mapping for each
+** page since the Virtual Coherence Index has to be generated
+** and updated for each page.
+**
+** PAGE_SIZE could be greater than IOVP_SIZE. But not the inverse.
+*/
+#define IOVP_SIZE	PAGE_SIZE
+#define IOVP_SHIFT	PAGE_SHIFT
+#define IOVP_MASK	PAGE_MASK
+
+#define SBA_PERF_CFG	0x708	/* Performance Counter stuff */
+#define SBA_PERF_MASK1	0x718
+#define SBA_PERF_MASK2	0x730
+
+
+/*
+** Offsets into PCI Performance Counters (functions 12 and 13)
+** Controlled by PERF registers in function 2 & 3 respectively.
+*/
+#define SBA_PERF_CNT1	0x200
+#define SBA_PERF_CNT2	0x208
+#define SBA_PERF_CNT3	0x210
+
+
+struct ioc {
+	void __iomem	*ioc_hpa;	/* I/O MMU base address */
+	char		*res_map;	/* resource map, bit == pdir entry */
+	u64		*pdir_base;	/* physical base address */
+	unsigned long	ibase;	/* pdir IOV Space base - shared w/lba_pci */
+	unsigned long	imask;	/* pdir IOV Space mask - shared w/lba_pci */
+#ifdef ZX1_SUPPORT
+	unsigned long	iovp_mask;	/* help convert IOVA to IOVP */
+#endif
+	unsigned long	*res_hint;	/* next avail IOVP - circular search */
+	spinlock_t	res_lock;
+	unsigned int	res_bitshift;	/* from the LEFT! */
+	unsigned int	res_size;	/* size of resource map in bytes */
+#if SBA_HINT_SUPPORT
+/* FIXME : DMA HINTs not used */
+	unsigned long	hint_mask_pdir;	/* bits used for DMA hints */
+	unsigned int	hint_shift_pdir;
+#endif
+#if DELAYED_RESOURCE_CNT > 0
+	int saved_cnt;
+	struct sba_dma_pair {
+		dma_addr_t	iova;
+		size_t		size;
+	} saved[DELAYED_RESOURCE_CNT];
+#endif
+
+#ifdef SBA_COLLECT_STATS
+#define SBA_SEARCH_SAMPLE	0x100
+	unsigned long avg_search[SBA_SEARCH_SAMPLE];
+	unsigned long avg_idx;	/* current index into avg_search */
+	unsigned long used_pages;
+	unsigned long msingle_calls;
+	unsigned long msingle_pages;
+	unsigned long msg_calls;
+	unsigned long msg_pages;
+	unsigned long usingle_calls;
+	unsigned long usingle_pages;
+	unsigned long usg_calls;
+	unsigned long usg_pages;
+#endif
+
+	/* STUFF We don't need in performance path */
+	unsigned int	pdir_size;	/* in bytes, determined by IOV Space size */
+};
+
+struct sba_device {
+	struct sba_device	*next;	/* list of SBA's in system */
+	struct parisc_device	*dev;	/* dev found in bus walk */
+	struct parisc_device_id	*iodc;	/* data about dev from firmware */
+	const char 		*name;
+	void __iomem		*sba_hpa; /* base address */
+	spinlock_t		sba_lock;
+	unsigned int		flags;  /* state/functionality enabled */
+	unsigned int		hw_rev;  /* HW revision of chip */
+
+	struct resource		chip_resv; /* MMIO reserved for chip */
+	struct resource		iommu_resv; /* MMIO reserved for iommu */
+
+	unsigned int		num_ioc;  /* number of on-board IOC's */
+	struct ioc		ioc[MAX_IOC];
+};
+
+
+static struct sba_device *sba_list;
+
+static unsigned long ioc_needs_fdc = 0;
+
+/* global count of IOMMUs in the system */
+static unsigned int global_ioc_cnt = 0;
+
+/* PA8700 (Piranha 2.2) bug workaround */
+static unsigned long piranha_bad_128k = 0;
+
+/* Looks nice and keeps the compiler happy */
+#define SBA_DEV(d) ((struct sba_device *) (d))
+
+#if SBA_AGP_SUPPORT
+static int reserve_sba_gart = 1;
+#endif
+
+#define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
+
+
+/************************************
+** SBA register read and write support
+**
+** BE WARNED: register writes are posted.
+**  (ie follow writes which must reach HW with a read)
+**
+** Superdome (in particular, REO) allows only 64-bit CSR accesses.
+*/
+#define READ_REG32(addr)	 le32_to_cpu(__raw_readl(addr))
+#define READ_REG64(addr)	 le64_to_cpu(__raw_readq(addr))
+#define WRITE_REG32(val, addr) __raw_writel(cpu_to_le32(val), addr)
+#define WRITE_REG64(val, addr) __raw_writeq(cpu_to_le64(val), addr)
+
+#ifdef __LP64__
+#define READ_REG(addr)		READ_REG64(addr)
+#define WRITE_REG(value, addr)	WRITE_REG64(value, addr)
+#else
+#define READ_REG(addr)		READ_REG32(addr)
+#define WRITE_REG(value, addr)	WRITE_REG32(value, addr)
+#endif
+
+#ifdef DEBUG_SBA_INIT
+
+/* NOTE: When __LP64__ isn't defined, READ_REG64() is two 32-bit reads */
+
+/**
+ * sba_dump_ranges - debugging only - print ranges assigned to this IOA
+ * @hpa: base address of the sba
+ *
+ * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
+ * IO Adapter (aka Bus Converter).
+ */
+static void
+sba_dump_ranges(void __iomem *hpa)
+{
+	DBG_INIT("SBA at 0x%p\n", hpa);
+	DBG_INIT("IOS_DIST_BASE   : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
+	DBG_INIT("IOS_DIST_MASK   : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
+	DBG_INIT("IOS_DIST_ROUTE  : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
+	DBG_INIT("\n");
+	DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
+	DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
+	DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
+}
+
+/**
+ * sba_dump_tlb - debugging only - print IOMMU operating parameters
+ * @hpa: base address of the IOMMU
+ *
+ * Print the size/location of the IO MMU PDIR.
+ */
+static void sba_dump_tlb(void __iomem *hpa)
+{
+	DBG_INIT("IO TLB at 0x%p\n", hpa);
+	DBG_INIT("IOC_IBASE    : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
+	DBG_INIT("IOC_IMASK    : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
+	DBG_INIT("IOC_TCNFG    : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
+	DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
+	DBG_INIT("\n");
+}
+#else
+#define sba_dump_ranges(x)
+#define sba_dump_tlb(x)
+#endif
+
+
+#ifdef ASSERT_PDIR_SANITY
+
+/**
+ * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @msg: text to print ont the output line.
+ * @pide: pdir index.
+ *
+ * Print one entry of the IO MMU PDIR in human readable form.
+ */
+static void
+sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
+{
+	/* start printing from lowest pde in rval */
+	u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
+	unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
+	uint rcnt;
+
+	printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
+		 msg,
+		 rptr, pide & (BITS_PER_LONG - 1), *rptr);
+
+	rcnt = 0;
+	while (rcnt < BITS_PER_LONG) {
+		printk(KERN_DEBUG "%s %2d %p %016Lx\n",
+			(rcnt == (pide & (BITS_PER_LONG - 1)))
+				? "    -->" : "       ",
+			rcnt, ptr, *ptr );
+		rcnt++;
+		ptr++;
+	}
+	printk(KERN_DEBUG "%s", msg);
+}
+
+
+/**
+ * sba_check_pdir - debugging only - consistency checker
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @msg: text to print ont the output line.
+ *
+ * Verify the resource map and pdir state is consistent
+ */
+static int
+sba_check_pdir(struct ioc *ioc, char *msg)
+{
+	u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
+	u32 *rptr = (u32 *) ioc->res_map;	/* resource map ptr */
+	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
+	uint pide = 0;
+
+	while (rptr < rptr_end) {
+		u32 rval = *rptr;
+		int rcnt = 32;	/* number of bits we might check */
+
+		while (rcnt) {
+			/* Get last byte and highest bit from that */
+			u32 pde = ((u32) (((char *)pptr)[7])) << 24;
+			if ((rval ^ pde) & 0x80000000)
+			{
+				/*
+				** BUMMER!  -- res_map != pdir --
+				** Dump rval and matching pdir entries
+				*/
+				sba_dump_pdir_entry(ioc, msg, pide);
+				return(1);
+			}
+			rcnt--;
+			rval <<= 1;	/* try the next bit */
+			pptr++;
+			pide++;
+		}
+		rptr++;	/* look at next word of res_map */
+	}
+	/* It'd be nice if we always got here :^) */
+	return 0;
+}
+
+
+/**
+ * sba_dump_sg - debugging only - print Scatter-Gather list
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @startsg: head of the SG list
+ * @nents: number of entries in SG list
+ *
+ * print the SG list so we can verify it's correct by hand.
+ */
+static void
+sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
+{
+	while (nents-- > 0) {
+		printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
+				nents,
+				(unsigned long) sg_dma_address(startsg),
+				sg_dma_len(startsg),
+				sg_virt_addr(startsg), startsg->length);
+		startsg++;
+	}
+}
+
+#endif /* ASSERT_PDIR_SANITY */
+
+
+
+
+/**************************************************************
+*
+*   I/O Pdir Resource Management
+*
+*   Bits set in the resource map are in use.
+*   Each bit can represent a number of pages.
+*   LSbs represent lower addresses (IOVA's).
+*
+***************************************************************/
+#define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
+
+/* Convert from IOVP to IOVA and vice versa. */
+
+#ifdef ZX1_SUPPORT
+/* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
+#define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
+#define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
+#else
+/* only support Astro and ancestors. Saves a few cycles in key places */
+#define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
+#define SBA_IOVP(ioc,iova) (iova)
+#endif
+
+#define PDIR_INDEX(iovp)   ((iovp)>>IOVP_SHIFT)
+
+#define RESMAP_MASK(n)    (~0UL << (BITS_PER_LONG - (n)))
+#define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)
+
+
+/**
+ * sba_search_bitmap - find free space in IO PDIR resource bitmap
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @bits_wanted: number of entries we need.
+ *
+ * Find consecutive free bits in resource bitmap.
+ * Each bit represents one entry in the IO Pdir.
+ * Cool perf optimization: search for log2(size) bits at a time.
+ */
+static SBA_INLINE unsigned long
+sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted)
+{
+	unsigned long *res_ptr = ioc->res_hint;
+	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
+	unsigned long pide = ~0UL;
+
+	if (bits_wanted > (BITS_PER_LONG/2)) {
+		/* Search word at a time - no mask needed */
+		for(; res_ptr < res_end; ++res_ptr) {
+			if (*res_ptr == 0) {
+				*res_ptr = RESMAP_MASK(bits_wanted);
+				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
+				pide <<= 3;	/* convert to bit address */
+				break;
+			}
+		}
+		/* point to the next word on next pass */
+		res_ptr++;
+		ioc->res_bitshift = 0;
+	} else {
+		/*
+		** Search the resource bit map on well-aligned values.
+		** "o" is the alignment.
+		** We need the alignment to invalidate I/O TLB using
+		** SBA HW features in the unmap path.
+		*/
+		unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
+		uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
+		unsigned long mask;
+
+		if (bitshiftcnt >= BITS_PER_LONG) {
+			bitshiftcnt = 0;
+			res_ptr++;
+		}
+		mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
+
+		DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
+		while(res_ptr < res_end)
+		{ 
+			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
+			WARN_ON(mask == 0);
+			if(((*res_ptr) & mask) == 0) {
+				*res_ptr |= mask;     /* mark resources busy! */
+				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
+				pide <<= 3;	/* convert to bit address */
+				pide += bitshiftcnt;
+				break;
+			}
+			mask >>= o;
+			bitshiftcnt += o;
+			if (mask == 0) {
+				mask = RESMAP_MASK(bits_wanted);
+				bitshiftcnt=0;
+				res_ptr++;
+			}
+		}
+		/* look in the same word on the next pass */
+		ioc->res_bitshift = bitshiftcnt + bits_wanted;
+	}
+
+	/* wrapped ? */
+	if (res_end <= res_ptr) {
+		ioc->res_hint = (unsigned long *) ioc->res_map;
+		ioc->res_bitshift = 0;
+	} else {
+		ioc->res_hint = res_ptr;
+	}
+	return (pide);
+}
+
+
+/**
+ * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @size: number of bytes to create a mapping for
+ *
+ * Given a size, find consecutive unmarked and then mark those bits in the
+ * resource bit map.
+ */
+static int
+sba_alloc_range(struct ioc *ioc, size_t size)
+{
+	unsigned int pages_needed = size >> IOVP_SHIFT;
+#ifdef SBA_COLLECT_STATS
+	unsigned long cr_start = mfctl(16);
+#endif
+	unsigned long pide;
+
+	pide = sba_search_bitmap(ioc, pages_needed);
+	if (pide >= (ioc->res_size << 3)) {
+		pide = sba_search_bitmap(ioc, pages_needed);
+		if (pide >= (ioc->res_size << 3))
+			panic("%s: I/O MMU @ %p is out of mapping resources\n",
+			      __FILE__, ioc->ioc_hpa);
+	}
+
+#ifdef ASSERT_PDIR_SANITY
+	/* verify the first enable bit is clear */
+	if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
+		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
+	}
+#endif
+
+	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
+		__FUNCTION__, size, pages_needed, pide,
+		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
+		ioc->res_bitshift );
+
+#ifdef SBA_COLLECT_STATS
+	{
+		unsigned long cr_end = mfctl(16);
+		unsigned long tmp = cr_end - cr_start;
+		/* check for roll over */
+		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
+	}
+	ioc->avg_search[ioc->avg_idx++] = cr_start;
+	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
+
+	ioc->used_pages += pages_needed;
+#endif
+
+	return (pide);
+}
+
+
+/**
+ * sba_free_range - unmark bits in IO PDIR resource bitmap
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @iova: IO virtual address which was previously allocated.
+ * @size: number of bytes to create a mapping for
+ *
+ * clear bits in the ioc's resource map
+ */
+static SBA_INLINE void
+sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
+{
+	unsigned long iovp = SBA_IOVP(ioc, iova);
+	unsigned int pide = PDIR_INDEX(iovp);
+	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
+	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
+
+	int bits_not_wanted = size >> IOVP_SHIFT;
+
+	/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
+	unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
+
+	DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
+		__FUNCTION__, (uint) iova, size,
+		bits_not_wanted, m, pide, res_ptr, *res_ptr);
+
+#ifdef SBA_COLLECT_STATS
+	ioc->used_pages -= bits_not_wanted;
+#endif
+
+	*res_ptr &= ~m;
+}
+
+
+/**************************************************************
+*
+*   "Dynamic DMA Mapping" support (aka "Coherent I/O")
+*
+***************************************************************/
+
+#if SBA_HINT_SUPPORT
+#define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
+#endif
+
+typedef unsigned long space_t;
+#define KERNEL_SPACE 0
+
+/**
+ * sba_io_pdir_entry - fill in one IO PDIR entry
+ * @pdir_ptr:  pointer to IO PDIR entry
+ * @sid: process Space ID - currently only support KERNEL_SPACE
+ * @vba: Virtual CPU address of buffer to map
+ * @hint: DMA hint set to use for this mapping
+ *
+ * SBA Mapping Routine
+ *
+ * Given a virtual address (vba, arg2) and space id, (sid, arg1)
+ * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
+ * pdir_ptr (arg0). 
+ * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
+ * for Astro/Ike looks like:
+ *
+ *
+ *  0                    19                                 51   55       63
+ * +-+---------------------+----------------------------------+----+--------+
+ * |V|        U            |            PPN[43:12]            | U  |   VI   |
+ * +-+---------------------+----------------------------------+----+--------+
+ *
+ * Pluto is basically identical, supports fewer physical address bits:
+ *
+ *  0                       23                              51   55       63
+ * +-+------------------------+-------------------------------+----+--------+
+ * |V|        U               |         PPN[39:12]            | U  |   VI   |
+ * +-+------------------------+-------------------------------+----+--------+
+ *
+ *  V  == Valid Bit  (Most Significant Bit is bit 0)
+ *  U  == Unused
+ * PPN == Physical Page Number
+ * VI  == Virtual Index (aka Coherent Index)
+ *
+ * LPA instruction output is put into PPN field.
+ * LCI (Load Coherence Index) instruction provides the "VI" bits.
+ *
+ * We pre-swap the bytes since PCX-W is Big Endian and the
+ * IOMMU uses little endian for the pdir.
+ */
+
+void SBA_INLINE
+sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
+		  unsigned long hint)
+{
+	u64 pa; /* physical address */
+	register unsigned ci; /* coherent index */
+
+	pa = virt_to_phys(vba);
+	pa &= IOVP_MASK;
+
+	mtsp(sid,1);
+	asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
+	pa |= (ci >> 12) & 0xff;  /* move CI (8 bits) into lowest byte */
+
+	pa |= 0x8000000000000000ULL;	/* set "valid" bit */
+	*pdir_ptr = cpu_to_le64(pa);	/* swap and store into I/O Pdir */
+
+	/*
+	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
+	 * (bit #61, big endian), we have to flush and sync every time
+	 * IO-PDIR is changed in Ike/Astro.
+	 */
+	if (ioc_needs_fdc) {
+		asm volatile("fdc 0(%%sr1,%0)\n\tsync" : : "r" (pdir_ptr));
+	}
+}
+
+
+/**
+ * sba_mark_invalid - invalidate one or more IO PDIR entries
+ * @ioc: IO MMU structure which owns the pdir we are interested in.
+ * @iova:  IO Virtual Address mapped earlier
+ * @byte_cnt:  number of bytes this mapping covers.
+ *
+ * Marking the IO PDIR entry(ies) as Invalid and invalidate
+ * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
+ * is to purge stale entries in the IO TLB when unmapping entries.
+ *
+ * The PCOM register supports purging of multiple pages, with a minium
+ * of 1 page and a maximum of 2GB. Hardware requires the address be
+ * aligned to the size of the range being purged. The size of the range
+ * must be a power of 2. The "Cool perf optimization" in the
+ * allocation routine helps keep that true.
+ */
+static SBA_INLINE void
+sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
+{
+	u32 iovp = (u32) SBA_IOVP(ioc,iova);
+
+	/* Even though this is a big-endian machine, the entries
+	** in the iopdir are little endian. That's why we clear the byte
+	** at +7 instead of at +0.
+	*/
+	int off = PDIR_INDEX(iovp)*sizeof(u64)+7;
+
+#ifdef ASSERT_PDIR_SANITY
+	/* Assert first pdir entry is set */
+	if (0x80 != (((u8 *) ioc->pdir_base)[off])) {
+		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
+	}
+#endif
+
+	if (byte_cnt <= IOVP_SIZE)
+	{
+		iovp |= IOVP_SHIFT;     /* set "size" field for PCOM */
+
+		/*
+		** clear I/O PDIR entry "valid" bit
+		** Do NOT clear the rest - save it for debugging.
+		** We should only clear bits that have previously
+		** been enabled.
+		*/
+		((u8 *)(ioc->pdir_base))[off] = 0;
+	} else {
+		u32 t = get_order(byte_cnt) + PAGE_SHIFT;
+
+		iovp |= t;
+		do {
+			/* clear I/O Pdir entry "valid" bit first */
+			((u8 *)(ioc->pdir_base))[off] = 0;
+			off += sizeof(u64);
+			byte_cnt -= IOVP_SIZE;
+		} while (byte_cnt > 0);
+	}
+
+	WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
+}
+
+/**
+ * sba_dma_supported - PCI driver can query DMA support
+ * @dev: instance of PCI owned by the driver that's asking
+ * @mask:  number of address bits this PCI device can handle
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static int sba_dma_supported( struct device *dev, u64 mask)
+{
+	struct ioc *ioc;
+	if (dev == NULL) {
+		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
+		BUG();
+		return(0);
+	}
+
+	ioc = GET_IOC(dev);
+
+	/* check if mask is > than the largest IO Virt Address */
+
+	return((int) (mask >= (ioc->ibase +
+				(ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
+}
+
+
+/**
+ * sba_map_single - map one buffer and return IOVA for DMA
+ * @dev: instance of PCI owned by the driver that's asking.
+ * @addr:  driver buffer to map.
+ * @size:  number of bytes to map in driver buffer.
+ * @direction:  R/W or both.
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static dma_addr_t
+sba_map_single(struct device *dev, void *addr, size_t size,
+	       enum dma_data_direction direction)
+{
+	struct ioc *ioc;
+	unsigned long flags; 
+	dma_addr_t iovp;
+	dma_addr_t offset;
+	u64 *pdir_start;
+	int pide;
+
+	ioc = GET_IOC(dev);
+
+	/* save offset bits */
+	offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
+
+	/* round up to nearest IOVP_SIZE */
+	size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
+
+	spin_lock_irqsave(&ioc->res_lock, flags);
+#ifdef ASSERT_PDIR_SANITY
+	sba_check_pdir(ioc,"Check before sba_map_single()");
+#endif
+
+#ifdef SBA_COLLECT_STATS
+	ioc->msingle_calls++;
+	ioc->msingle_pages += size >> IOVP_SHIFT;
+#endif
+	pide = sba_alloc_range(ioc, size);
+	iovp = (dma_addr_t) pide << IOVP_SHIFT;
+
+	DBG_RUN("%s() 0x%p -> 0x%lx\n",
+		__FUNCTION__, addr, (long) iovp | offset);
+
+	pdir_start = &(ioc->pdir_base[pide]);
+
+	while (size > 0) {
+		sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
+
+		DBG_RUN("	pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
+			pdir_start,
+			(u8) (((u8 *) pdir_start)[7]),
+			(u8) (((u8 *) pdir_start)[6]),
+			(u8) (((u8 *) pdir_start)[5]),
+			(u8) (((u8 *) pdir_start)[4]),
+			(u8) (((u8 *) pdir_start)[3]),
+			(u8) (((u8 *) pdir_start)[2]),
+			(u8) (((u8 *) pdir_start)[1]),
+			(u8) (((u8 *) pdir_start)[0])
+			);
+
+		addr += IOVP_SIZE;
+		size -= IOVP_SIZE;
+		pdir_start++;
+	}
+	/* form complete address */
+#ifdef ASSERT_PDIR_SANITY
+	sba_check_pdir(ioc,"Check after sba_map_single()");
+#endif
+	spin_unlock_irqrestore(&ioc->res_lock, flags);
+	return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
+}
+
+
+/**
+ * sba_unmap_single - unmap one IOVA and free resources
+ * @dev: instance of PCI owned by the driver that's asking.
+ * @iova:  IOVA of driver buffer previously mapped.
+ * @size:  number of bytes mapped in driver buffer.
+ * @direction:  R/W or both.
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static void
+sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size,
+		 enum dma_data_direction direction)
+{
+	struct ioc *ioc;
+#if DELAYED_RESOURCE_CNT > 0
+	struct sba_dma_pair *d;
+#endif
+	unsigned long flags; 
+	dma_addr_t offset;
+
+	DBG_RUN("%s() iovp 0x%lx/%x\n", __FUNCTION__, (long) iova, size);
+
+	ioc = GET_IOC(dev);
+	offset = iova & ~IOVP_MASK;
+	iova ^= offset;        /* clear offset bits */
+	size += offset;
+	size = ROUNDUP(size, IOVP_SIZE);
+
+	spin_lock_irqsave(&ioc->res_lock, flags);
+
+#ifdef SBA_COLLECT_STATS
+	ioc->usingle_calls++;
+	ioc->usingle_pages += size >> IOVP_SHIFT;
+#endif
+
+	sba_mark_invalid(ioc, iova, size);
+
+#if DELAYED_RESOURCE_CNT > 0
+	/* Delaying when we re-use a IO Pdir entry reduces the number
+	 * of MMIO reads needed to flush writes to the PCOM register.
+	 */
+	d = &(ioc->saved[ioc->saved_cnt]);
+	d->iova = iova;
+	d->size = size;
+	if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
+		int cnt = ioc->saved_cnt;
+		while (cnt--) {
+			sba_free_range(ioc, d->iova, d->size);
+			d--;
+		}
+		ioc->saved_cnt = 0;
+		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
+	}
+#else /* DELAYED_RESOURCE_CNT == 0 */
+	sba_free_range(ioc, iova, size);
+	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
+#endif /* DELAYED_RESOURCE_CNT == 0 */
+	spin_unlock_irqrestore(&ioc->res_lock, flags);
+
+	/* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
+	** For Astro based systems this isn't a big deal WRT performance.
+	** As long as 2.4 kernels copyin/copyout data from/to userspace,
+	** we don't need the syncdma. The issue here is I/O MMU cachelines
+	** are *not* coherent in all cases.  May be hwrev dependent.
+	** Need to investigate more.
+	asm volatile("syncdma");	
+	*/
+}
+
+
+/**
+ * sba_alloc_consistent - allocate/map shared mem for DMA
+ * @hwdev: instance of PCI owned by the driver that's asking.
+ * @size:  number of bytes mapped in driver buffer.
+ * @dma_handle:  IOVA of new buffer.
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static void *sba_alloc_consistent(struct device *hwdev, size_t size,
+					dma_addr_t *dma_handle, int gfp)
+{
+	void *ret;
+
+	if (!hwdev) {
+		/* only support PCI */
+		*dma_handle = 0;
+		return 0;
+	}
+
+        ret = (void *) __get_free_pages(gfp, get_order(size));
+
+	if (ret) {
+		memset(ret, 0, size);
+		*dma_handle = sba_map_single(hwdev, ret, size, 0);
+	}
+
+	return ret;
+}
+
+
+/**
+ * sba_free_consistent - free/unmap shared mem for DMA
+ * @hwdev: instance of PCI owned by the driver that's asking.
+ * @size:  number of bytes mapped in driver buffer.
+ * @vaddr:  virtual address IOVA of "consistent" buffer.
+ * @dma_handler:  IO virtual address of "consistent" buffer.
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static void
+sba_free_consistent(struct device *hwdev, size_t size, void *vaddr,
+		    dma_addr_t dma_handle)
+{
+	sba_unmap_single(hwdev, dma_handle, size, 0);
+	free_pages((unsigned long) vaddr, get_order(size));
+}
+
+
+/*
+** Since 0 is a valid pdir_base index value, can't use that
+** to determine if a value is valid or not. Use a flag to indicate
+** the SG list entry contains a valid pdir index.
+*/
+#define PIDE_FLAG 0x80000000UL
+
+#ifdef SBA_COLLECT_STATS
+#define IOMMU_MAP_STATS
+#endif
+#include "iommu-helpers.h"
+
+#ifdef DEBUG_LARGE_SG_ENTRIES
+int dump_run_sg = 0;
+#endif
+
+
+/**
+ * sba_map_sg - map Scatter/Gather list
+ * @dev: instance of PCI owned by the driver that's asking.
+ * @sglist:  array of buffer/length pairs
+ * @nents:  number of entries in list
+ * @direction:  R/W or both.
+ *
+ * See Documentation/DMA-mapping.txt
+ */
+static int
+sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
+	   enum dma_data_direction direction)
+{
+	struct ioc *ioc;
+	int coalesced, filled = 0;
+	unsigned long flags;
+
+	DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents);
+
+	ioc = GET_IOC(dev);
+
+	/* Fast path single entry scatterlists. */
+	if (nents == 1) {
+		sg_dma_address(sglist) = sba_map_single(dev,
+						(void *)sg_virt_addr(sglist),
+						sglist->length, direction);
+		sg_dma_len(sglist)     = sglist->length;
+		return 1;
+	}
+
+	spin_lock_irqsave(&ioc->res_lock, flags);
+
+#ifdef ASSERT_PDIR_SANITY
+	if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
+	{
+		sba_dump_sg(ioc, sglist, nents);
+		panic("Check before sba_map_sg()");
+	}
+#endif
+
+#ifdef SBA_COLLECT_STATS
+	ioc->msg_calls++;
+#endif
+
+	/*
+	** First coalesce the chunks and allocate I/O pdir space
+	**
+	** If this is one DMA stream, we can properly map using the
+	** correct virtual address associated with each DMA page.
+	** w/o this association, we wouldn't have coherent DMA!
+	** Access to the virtual address is what forces a two pass algorithm.
+	*/
+	coalesced = iommu_coalesce_chunks(ioc, sglist, nents, sba_alloc_range);
+
+	/*
+	** Program the I/O Pdir
+	**
+	** map the virtual addresses to the I/O Pdir
+	** o dma_address will contain the pdir index
+	** o dma_len will contain the number of bytes to map 
+	** o address contains the virtual address.
+	*/
+	filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
+
+#ifdef ASSERT_PDIR_SANITY
+	if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
+	{
+		sba_dump_sg(ioc, sglist, nents);
+		panic("Check after sba_map_sg()\n");
+	}
+#endif
+
+	spin_unlock_irqrestore(&ioc->res_lock, flags);
+
+	DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled);
+
+	return filled;
+}
+
+
+/**
+ * sba_unmap_sg - unmap Scatter/Gather list
+ * @dev: instance of PCI owned by the driver that's asking.
+ * @sglist:  array of buffer/length pairs