amd64-agp.c

优龙2410linux2.6.8内核源代码

/*
 * Copyright 2001-2003 SuSE Labs.
 * Distributed under the GNU public license, v2.
 *
 * This is a GART driver for the AMD Opteron/Athlon64 on-CPU northbridge.
 * It also includes support for the AMD 8151 AGP bridge,
 * although it doesn't actually do much, as all the real
 * work is done in the northbridge(s).
 */

#include <linux/config.h>
#include <linux/module.h>
#include <linux/pci.h>
#include <linux/init.h>
#include <linux/agp_backend.h>
#include "agp.h"

/* Will need to be increased if AMD64 ever goes >8-way. */
#define MAX_HAMMER_GARTS   8

/* PTE bits. */
#define GPTE_VALID	1
#define GPTE_COHERENT	2

/* Aperture control register bits. */
#define GARTEN		(1<<0)
#define DISGARTCPU	(1<<4)
#define DISGARTIO	(1<<5)

/* GART cache control register bits. */
#define INVGART		(1<<0)
#define GARTPTEERR	(1<<1)

/* K8 On-cpu GART registers */
#define AMD64_GARTAPERTURECTL	0x90
#define AMD64_GARTAPERTUREBASE	0x94
#define AMD64_GARTTABLEBASE	0x98
#define AMD64_GARTCACHECTL	0x9c
#define AMD64_GARTEN		(1<<0)

/* NVIDIA K8 registers */
#define NVIDIA_X86_64_0_APBASE		0x10
#define NVIDIA_X86_64_1_APBASE1		0x50
#define NVIDIA_X86_64_1_APLIMIT1	0x54
#define NVIDIA_X86_64_1_APSIZE		0xa8
#define NVIDIA_X86_64_1_APBASE2		0xd8
#define NVIDIA_X86_64_1_APLIMIT2	0xdc

static int nr_garts;
static struct pci_dev * hammers[MAX_HAMMER_GARTS];

static struct resource *aperture_resource;
static int __initdata agp_try_unsupported;

static int gart_iterator;
#define for_each_nb() for(gart_iterator=0;gart_iterator<nr_garts;gart_iterator++)

static void flush_amd64_tlb(struct pci_dev *dev)
{
	u32 tmp;

	pci_read_config_dword (dev, AMD64_GARTCACHECTL, &tmp);
	tmp |= INVGART;
	pci_write_config_dword (dev, AMD64_GARTCACHECTL, tmp);
}

static void amd64_tlbflush(struct agp_memory *temp)
{
	for_each_nb()
		flush_amd64_tlb(hammers[gart_iterator]);
}

static int amd64_insert_memory(struct agp_memory *mem, off_t pg_start, int type)
{
	int i, j, num_entries;
	long tmp;
	u32 pte;

	num_entries = agp_num_entries();

	if (type != 0 || mem->type != 0)
		return -EINVAL;

	/* Make sure we can fit the range in the gatt table. */
	/* FIXME: could wrap */
	if (((unsigned long)pg_start + mem->page_count) > num_entries)
		return -EINVAL;

	j = pg_start;

	/* gatt table should be empty. */
	while (j < (pg_start + mem->page_count)) {
		if (!PGE_EMPTY(agp_bridge, agp_bridge->gatt_table[j]))
			return -EBUSY;
		j++;
	}

	if (mem->is_flushed == FALSE) {
		global_cache_flush();
		mem->is_flushed = TRUE;
	}

	for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
		tmp = agp_bridge->driver->mask_memory(mem->memory[i], mem->type);

		BUG_ON(tmp & 0xffffff0000000ffcULL);
		pte = (tmp & 0x000000ff00000000ULL) >> 28;
		pte |=(tmp & 0x00000000fffff000ULL);
		pte |= GPTE_VALID | GPTE_COHERENT;

		agp_bridge->gatt_table[j] = pte;
	}
	amd64_tlbflush(mem);
	return 0;
}

/*
 * This hack alters the order element according
 * to the size of a long. It sucks. I totally disown this, even
 * though it does appear to work for the most part.
 */
static struct aper_size_info_32 amd64_aperture_sizes[7] =
{
	{32,   8192,   3+(sizeof(long)/8), 0 },
	{64,   16384,  4+(sizeof(long)/8), 1<<1 },
	{128,  32768,  5+(sizeof(long)/8), 1<<2 },
	{256,  65536,  6+(sizeof(long)/8), 1<<1 | 1<<2 },
	{512,  131072, 7+(sizeof(long)/8), 1<<3 },
	{1024, 262144, 8+(sizeof(long)/8), 1<<1 | 1<<3},
	{2048, 524288, 9+(sizeof(long)/8), 1<<2 | 1<<3}
};


/*
 * Get the current Aperture size from the x86-64.
 * Note, that there may be multiple x86-64's, but we just return
 * the value from the first one we find. The set_size functions
 * keep the rest coherent anyway. Or at least should do.
 */
static int amd64_fetch_size(void)
{
	struct pci_dev *dev;
	int i;
	u32 temp;
	struct aper_size_info_32 *values;

	dev = hammers[0];
	if (dev==NULL)
		return 0;

	pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &temp);
	temp = (temp & 0xe);
	values = A_SIZE_32(amd64_aperture_sizes);

	for (i = 0; i < agp_bridge->driver->num_aperture_sizes; i++) {
		if (temp == values[i].size_value) {
			agp_bridge->previous_size =
			    agp_bridge->current_size = (void *) (values + i);

			agp_bridge->aperture_size_idx = i;
			return values[i].size;
		}
	}
	return 0;
}

/*
 * In a multiprocessor x86-64 system, this function gets
 * called once for each CPU.
 */
static u64 amd64_configure (struct pci_dev *hammer, u64 gatt_table)
{
	u64 aperturebase;
	u32 tmp;
	u64 addr, aper_base;

	/* Address to map to */
	pci_read_config_dword (hammer, AMD64_GARTAPERTUREBASE, &tmp);
	aperturebase = tmp << 25;
	aper_base = (aperturebase & PCI_BASE_ADDRESS_MEM_MASK);

	/* address of the mappings table */
	addr = (u64) gatt_table;
	addr >>= 12;
	tmp = (u32) addr<<4;
	tmp &= ~0xf;
	pci_write_config_dword (hammer, AMD64_GARTTABLEBASE, tmp);

	/* Enable GART translation for this hammer. */
	pci_read_config_dword(hammer, AMD64_GARTAPERTURECTL, &tmp);
	tmp |= GARTEN;
	tmp &= ~(DISGARTCPU | DISGARTIO);
	pci_write_config_dword(hammer, AMD64_GARTAPERTURECTL, tmp);

	/* keep CPU's coherent. */
	flush_amd64_tlb (hammer);

	return aper_base;
}


static struct aper_size_info_32 amd_8151_sizes[7] =
{
	{2048, 524288, 9, 0x00000000 },	/* 0 0 0 0 0 0 */
	{1024, 262144, 8, 0x00000400 },	/* 1 0 0 0 0 0 */
	{512,  131072, 7, 0x00000600 },	/* 1 1 0 0 0 0 */
	{256,  65536,  6, 0x00000700 },	/* 1 1 1 0 0 0 */
	{128,  32768,  5, 0x00000720 },	/* 1 1 1 1 0 0 */
	{64,   16384,  4, 0x00000730 },	/* 1 1 1 1 1 0 */
	{32,   8192,   3, 0x00000738 } 	/* 1 1 1 1 1 1 */
};

static int amd_8151_configure(void)
{
	unsigned long gatt_bus = virt_to_phys(agp_bridge->gatt_table_real);

	/* Configure AGP regs in each x86-64 host bridge. */
	for_each_nb() {
		agp_bridge->gart_bus_addr =
				amd64_configure(hammers[gart_iterator],gatt_bus);
	}
	return 0;
}


static void amd64_cleanup(void)
{
	u32 tmp;

	for_each_nb() {
		/* disable gart translation */
		pci_read_config_dword (hammers[gart_iterator], AMD64_GARTAPERTURECTL, &tmp);
		tmp &= ~AMD64_GARTEN;
		pci_write_config_dword (hammers[gart_iterator], AMD64_GARTAPERTURECTL, tmp);
	}
}


struct agp_bridge_driver amd_8151_driver = {
	.owner			= THIS_MODULE,
	.aperture_sizes		= amd_8151_sizes,
	.size_type		= U32_APER_SIZE,
	.num_aperture_sizes	= 7,
	.configure		= amd_8151_configure,
	.fetch_size		= amd64_fetch_size,
	.cleanup		= amd64_cleanup,
	.tlb_flush		= amd64_tlbflush,
	.mask_memory		= agp_generic_mask_memory,
	.masks			= NULL,
	.agp_enable		= agp_generic_enable,
	.cache_flush		= global_cache_flush,
	.create_gatt_table	= agp_generic_create_gatt_table,
	.free_gatt_table	= agp_generic_free_gatt_table,
	.insert_memory		= amd64_insert_memory,
	.remove_memory		= agp_generic_remove_memory,
	.alloc_by_type		= agp_generic_alloc_by_type,
	.free_by_type		= agp_generic_free_by_type,
	.agp_alloc_page		= agp_generic_alloc_page,
	.agp_destroy_page	= agp_generic_destroy_page,
};

/* Some basic sanity checks for the aperture. */
static int __devinit aperture_valid(u64 aper, u32 size)
{
	u32 pfn, c;
	if (aper == 0) {
		printk(KERN_ERR PFX "No aperture\n");
		return 0;
	}
	if (size < 32*1024*1024) {
		printk(KERN_ERR PFX "Aperture too small (%d MB)\n", size>>20);
		return 0;
	}
	if (aper + size > 0xffffffff) {
		printk(KERN_ERR PFX "Aperture out of bounds\n");
		return 0;
	}
	pfn = aper >> PAGE_SHIFT;
	for (c = 0; c < size/PAGE_SIZE; c++) {
		if (!pfn_valid(pfn + c))
			break;
		if (!PageReserved(pfn_to_page(pfn + c))) {
			printk(KERN_ERR PFX "Aperture pointing to RAM\n");
			return 0;
		}
	}

	/* Request the Aperture. This catches cases when someone else
	   already put a mapping in there - happens with some very broken BIOS

	   Maybe better to use pci_assign_resource/pci_enable_device instead
	   trusting the bridges? */
	if (!aperture_resource &&
	    !(aperture_resource = request_mem_region(aper, size, "aperture"))) {
		printk(KERN_ERR PFX "Aperture conflicts with PCI mapping.\n");
		return 0;
	}
	return 1;
}

/*
 * W*s centric BIOS sometimes only set up the aperture in the AGP
 * bridge, not the northbridge. On AMD64 this is handled early
 * in aperture.c, but when GART_IOMMU is not enabled or we run
 * on a 32bit kernel this needs to be redone.
 * Unfortunately it is impossible to fix the aperture here because it's too late
 * to allocate that much memory. But at least error out cleanly instead of
 * crashing.
 */
static __devinit int fix_northbridge(struct pci_dev *nb, struct pci_dev *agp,
								 u16 cap)
{
	u32 aper_low, aper_hi;
	u64 aper, nb_aper;
	int order = 0;
	u32 nb_order, nb_base;
	u16 apsize;

	pci_read_config_dword(nb, 0x90, &nb_order);
	nb_order = (nb_order >> 1) & 7;
	pci_read_config_dword(nb, 0x94, &nb_base);
	nb_aper = nb_base << 25;
	if (aperture_valid(nb_aper, (32*1024*1024)<<nb_order)) {
		return 0;
	}

	/* Northbridge seems to contain crap. Try the AGP bridge. */

	pci_read_config_word(agp, cap+0x14, &apsize);
	if (apsize == 0xffff)
		return -1;

	apsize &= 0xfff;
	/* Some BIOS use weird encodings not in the AGPv3 table. */
	if (apsize & 0xff)
		apsize |= 0xf00;
	order = 7 - hweight16(apsize);

	pci_read_config_dword(agp, 0x10, &aper_low);
	pci_read_config_dword(agp, 0x14, &aper_hi);