sba_iommu.c

linux-2.6.15.6

/*
**  IA64 System Bus Adapter (SBA) I/O MMU manager
**
**	(c) Copyright 2002-2005 Alex Williamson
**	(c) Copyright 2002-2003 Grant Grundler
**	(c) Copyright 2002-2005 Hewlett-Packard Company
**
**	Portions (c) 2000 Grant Grundler (from parisc I/O MMU code)
**	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 HP
** McKinley machines and their successors.
**
*/

#include <linux/config.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/module.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 <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/acpi.h>
#include <linux/efi.h>
#include <linux/nodemask.h>
#include <linux/bitops.h>         /* hweight64() */

#include <asm/delay.h>		/* ia64_get_itc() */
#include <asm/io.h>
#include <asm/page.h>		/* PAGE_OFFSET */
#include <asm/dma.h>
#include <asm/system.h>		/* wmb() */

#include <asm/acpi-ext.h>

#define PFX "IOC: "

/*
** Enabling timing search of the pdir resource map.  Output in /proc.
** Disabled by default to optimize performance.
*/
#undef PDIR_SEARCH_TIMING

/*
** This option allows cards capable of 64bit DMA to bypass the IOMMU.  If
** not defined, all DMA will be 32bit and go through the TLB.
** There's potentially a conflict in the bio merge code with us
** advertising an iommu, but then bypassing it.  Since I/O MMU bypassing
** appears to give more performance than bio-level virtual merging, we'll
** do the former for now.  NOTE: BYPASS_SG also needs to be undef'd to
** completely restrict DMA to the IOMMU.
*/
#define ALLOW_IOV_BYPASS

/*
** This option specifically allows/disallows bypassing scatterlists with
** multiple entries.  Coalescing these entries can allow better DMA streaming
** and in some cases shows better performance than entirely bypassing the
** IOMMU.  Performance increase on the order of 1-2% sequential output/input
** using bonnie++ on a RAID0 MD device (sym2 & mpt).
*/
#undef ALLOW_IOV_BYPASS_SG

/*
** If a device prefetches beyond the end of a valid pdir entry, it will cause
** a hard failure, ie. MCA.  Version 3.0 and later of the zx1 LBA should
** disconnect on 4k boundaries and prevent such issues.  If the device is
** particularly agressive, this option will keep the entire pdir valid such
** that prefetching will hit a valid address.  This could severely impact
** error containment, and is therefore off by default.  The page that is
** used for spill-over is poisoned, so that should help debugging somewhat.
*/
#undef FULL_VALID_PDIR

#define ENABLE_MARK_CLEAN

/*
** The number of debug flags is a clue - this code is fragile.  NOTE: since
** tightening the use of res_lock the resource bitmap and actual pdir are no
** longer guaranteed to stay in sync.  The sanity checking code isn't going to
** like that.
*/
#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_BYPASS

#if defined(FULL_VALID_PDIR) && defined(ASSERT_PDIR_SANITY)
#error FULL_VALID_PDIR and ASSERT_PDIR_SANITY are mutually exclusive
#endif

#define SBA_INLINE	__inline__
/* #define SBA_INLINE */

#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

#ifdef DEBUG_BYPASS
#define DBG_BYPASS(x...)	printk(x)
#else
#define DBG_BYPASS(x...)
#endif

#ifdef ASSERT_PDIR_SANITY
#define ASSERT(expr) \
        if(!(expr)) { \
                printk( "\n" __FILE__ ":%d: Assertion " #expr " failed!\n",__LINE__); \
                panic(#expr); \
        }
#else
#define ASSERT(expr)
#endif

/*
** The number of pdir entries to "free" before issuing
** 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	64

#define PCI_DEVICE_ID_HP_SX2000_IOC	0x12ec

#define ZX1_IOC_ID	((PCI_DEVICE_ID_HP_ZX1_IOC << 16) | PCI_VENDOR_ID_HP)
#define ZX2_IOC_ID	((PCI_DEVICE_ID_HP_ZX2_IOC << 16) | PCI_VENDOR_ID_HP)
#define REO_IOC_ID	((PCI_DEVICE_ID_HP_REO_IOC << 16) | PCI_VENDOR_ID_HP)
#define SX1000_IOC_ID	((PCI_DEVICE_ID_HP_SX1000_IOC << 16) | PCI_VENDOR_ID_HP)
#define SX2000_IOC_ID	((PCI_DEVICE_ID_HP_SX2000_IOC << 16) | PCI_VENDOR_ID_HP)

#define ZX1_IOC_OFFSET	0x1000	/* ACPI reports SBA, we want IOC */

#define IOC_FUNC_ID	0x000
#define IOC_FCLASS	0x008	/* function class, bist, header, rev... */
#define IOC_IBASE	0x300	/* IO TLB */
#define IOC_IMASK	0x308
#define IOC_PCOM	0x310
#define IOC_TCNFG	0x318
#define IOC_PDIR_BASE	0x320

#define IOC_ROPE0_CFG	0x500
#define   IOC_ROPE_AO	  0x10	/* Allow "Relaxed Ordering" */


/* AGP GART driver looks for this */
#define ZX1_SBA_IOMMU_COOKIE	0x0000badbadc0ffeeUL

/*
** The zx1 IOC supports 4/8/16/64KB page sizes (see TCNFG register)
**
** Some IOCs (sx1000) can run at the above pages sizes, but are
** really only supported using the IOC at a 4k page size.
**
** iovp_size could only be greater than PAGE_SIZE if we are
** confident the drivers really only touch the next physical
** page iff that driver instance owns it.
*/
static unsigned long iovp_size;
static unsigned long iovp_shift;
static unsigned long iovp_mask;

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 */
	unsigned long	imask;		/* pdir IOV Space mask */

	unsigned long	*res_hint;	/* next avail IOVP - circular search */
	unsigned long	dma_mask;
	spinlock_t	res_lock;	/* protects the resource bitmap, but must be held when */
					/* clearing pdir to prevent races with allocations. */
	unsigned int	res_bitshift;	/* from the RIGHT! */
	unsigned int	res_size;	/* size of resource map in bytes */
#ifdef CONFIG_NUMA
	unsigned int	node;		/* node where this IOC lives */
#endif
#if DELAYED_RESOURCE_CNT > 0
	spinlock_t	saved_lock;	/* may want to try to get this on a separate cacheline */
					/* than res_lock for bigger systems. */
	int		saved_cnt;
	struct sba_dma_pair {
		dma_addr_t	iova;
		size_t		size;
	} saved[DELAYED_RESOURCE_CNT];
#endif

#ifdef PDIR_SEARCH_TIMING
#define SBA_SEARCH_SAMPLE	0x100
	unsigned long avg_search[SBA_SEARCH_SAMPLE];
	unsigned long avg_idx;	/* current index into avg_search */
#endif

	/* Stuff we don't need in performance path */
	struct ioc	*next;		/* list of IOC's in system */
	acpi_handle	handle;		/* for multiple IOC's */
	const char 	*name;
	unsigned int	func_id;
	unsigned int	rev;		/* HW revision of chip */
	u32		iov_size;
	unsigned int	pdir_size;	/* in bytes, determined by IOV Space size */
	struct pci_dev	*sac_only_dev;
};

static struct ioc *ioc_list;
static int reserve_sba_gart = 1;

static SBA_INLINE void sba_mark_invalid(struct ioc *, dma_addr_t, size_t);
static SBA_INLINE void sba_free_range(struct ioc *, dma_addr_t, size_t);

#define sba_sg_address(sg)	(page_address((sg)->page) + (sg)->offset)

#ifdef FULL_VALID_PDIR
static u64 prefetch_spill_page;
#endif

#ifdef CONFIG_PCI
# define GET_IOC(dev)	(((dev)->bus == &pci_bus_type)						\
			 ? ((struct ioc *) PCI_CONTROLLER(to_pci_dev(dev))->iommu) : NULL)
#else
# define GET_IOC(dev)	NULL
#endif

/*
** DMA_CHUNK_SIZE is used by the SCSI mid-layer to break up
** (or rather not merge) DMA's into managable chunks.
** On parisc, this is more of the software/tuning constraint
** rather than the HW. I/O MMU allocation alogorithms can be
** faster with smaller size is (to some degree).
*/
#define DMA_CHUNK_SIZE  (BITS_PER_LONG*iovp_size)

#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)
**
*/
#define READ_REG(addr)       __raw_readq(addr)
#define WRITE_REG(val, addr) __raw_writeq(val, addr)

#ifdef DEBUG_SBA_INIT

/**
 * 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(char *hpa)
{
	DBG_INIT("IO TLB at 0x%p\n", (void *)hpa);
	DBG_INIT("IOC_IBASE    : %016lx\n", READ_REG(hpa+IOC_IBASE));
	DBG_INIT("IOC_IMASK    : %016lx\n", READ_REG(hpa+IOC_IMASK));
	DBG_INIT("IOC_TCNFG    : %016lx\n", READ_REG(hpa+IOC_TCNFG));
	DBG_INIT("IOC_PDIR_BASE: %016lx\n", READ_REG(hpa+IOC_PDIR_BASE));
	DBG_INIT("\n");
}
#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  & ~(BITS_PER_LONG - 1)];
	unsigned long *rptr = (unsigned long *) &ioc->res_map[(pide >>3) & -sizeof(unsigned long)];
	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, (unsigned long long) *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)
{
	u64 *rptr_end = (u64 *) &(ioc->res_map[ioc->res_size]);
	u64 *rptr = (u64 *) ioc->res_map;	/* resource map ptr */
	u64 *pptr = ioc->pdir_base;	/* pdir ptr */
	uint pide = 0;

	while (rptr < rptr_end) {
		u64 rval;
		int rcnt; /* number of bits we might check */

		rval = *rptr;
		rcnt = 64;

		while (rcnt) {
			/* Get last byte and highest bit from that */
			u32 pde = ((u32)((*pptr >> (63)) & 0x1));
			if ((rval & 0x1) ^ pde)
			{
				/*
				** 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 : DMA %08lx/%05x CPU %p\n", nents,
		       startsg->dma_address, startsg->dma_length,
		       sba_sg_address(startsg));
		startsg++;
	}
}

static void
sba_check_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
{
	struct scatterlist *the_sg = startsg;
	int the_nents = nents;

	while (the_nents-- > 0) {
		if (sba_sg_address(the_sg) == 0x0UL)
			sba_dump_sg(NULL, startsg, nents);
		the_sg++;
	}
}

#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. */
#define SBA_IOVA(ioc,iovp,offset) ((ioc->ibase) | (iovp) | (offset))
#define SBA_IOVP(ioc,iova) ((iova) & ~(ioc->ibase))

#define PDIR_ENTRY_SIZE	sizeof(u64)

#define PDIR_INDEX(iovp)   ((iovp)>>iovp_shift)

#define RESMAP_MASK(n)    ~(~0UL << (n))
#define RESMAP_IDX_MASK   (sizeof(unsigned long) - 1)


/**
 * For most cases the normal get_order is sufficient, however it limits us
 * to PAGE_SIZE being the minimum mapping alignment and TC flush granularity.
 * It only incurs about 1 clock cycle to use this one with the static variable
 * and makes the code more intuitive.
 */
static SBA_INLINE int
get_iovp_order (unsigned long size)
{
	long double d = size - 1;
	long order;

	order = ia64_getf_exp(d);
	order = order - iovp_shift - 0xffff + 1;
	if (order < 0)
		order = 0;
	return order;
}

/**
 * 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.
 * @use_hint: use res_hint to indicate where to start looking
 *
 * 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, int use_hint)
{
	unsigned long *res_ptr;
	unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
	unsigned long flags, pide = ~0UL;

	ASSERT(((unsigned long) ioc->res_hint & (sizeof(unsigned long) - 1UL)) == 0);
	ASSERT(res_ptr < res_end);

	spin_lock_irqsave(&ioc->res_lock, flags);

	/* Allow caller to force a search through the entire resource space */
	if (likely(use_hint)) {
		res_ptr = ioc->res_hint;
	} else {
		res_ptr = (ulong *)ioc->res_map;
		ioc->res_bitshift = 0;
	}

	/*
	 * N.B.  REO/Grande defect AR2305 can cause TLB fetch timeouts
	 * if a TLB entry is purged while in use.  sba_mark_invalid()
	 * purges IOTLB entries in power-of-two sizes, so we also
	 * allocate IOVA space in power-of-two sizes.
	 */
	bits_wanted = 1UL << get_iovp_order(bits_wanted << iovp_shift);

	if (likely(bits_wanted == 1)) {
		unsigned int bitshiftcnt;
		for(; res_ptr < res_end ; res_ptr++) {
			if (likely(*res_ptr != ~0UL)) {
				bitshiftcnt = ffz(*res_ptr);
				*res_ptr |= (1UL << bitshiftcnt);
				pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
				pide <<= 3;	/* convert to bit address */
				pide += bitshiftcnt;
				ioc->res_bitshift = bitshiftcnt + bits_wanted;
				goto found_it;
			}
		}
		goto not_found;

	}
	
	if (likely(bits_wanted <= BITS_PER_LONG/2)) {
		/*
		** 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_iovp_order(bits_wanted << iovp_shift);
		uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
		unsigned long mask, base_mask;

		base_mask = RESMAP_MASK(bits_wanted);
		mask = base_mask << bitshiftcnt;

		DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
		for(; res_ptr < res_end ; res_ptr++)
		{ 
			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
			ASSERT(0 != mask);