agp_amd.c

#ifdef CONFIG_AGP_AMD

typedef struct _amd_page_map {
	unsigned long *real;
	unsigned long *remapped;
} amd_page_map;

static struct _amd_irongate_private {
	volatile u8 *registers;
	amd_page_map **gatt_pages;
	int num_tables;
	unsigned long *real;
	unsigned long *remapped;
} amd_irongate_private;

static int amd_create_page_map(amd_page_map *page_map)
{
	int i;
	int err = 0;

	page_map->real = (unsigned long *) __get_free_page(GFP_KERNEL);
	if (page_map->real == NULL) {
		return -ENOMEM;
	}
	SetPageReserved(virt_to_page(page_map->real));
	CACHE_FLUSH();
#ifdef CONFIG_X86
	err = change_page_attr(virt_to_page(page_map->real), 1, PAGE_KERNEL_NOCACHE);
#endif
	if (!err) 
	page_map->remapped = ioremap_nocache(virt_to_phys(page_map->real), 
					    PAGE_SIZE);
	if (page_map->remapped == NULL || err) {
		ClearPageReserved(virt_to_page(page_map->real));
		free_page((unsigned long) page_map->real);
		page_map->real = NULL;
		return -ENOMEM;
	}
	CACHE_FLUSH();

	for(i = 0; i < PAGE_SIZE / sizeof(unsigned long); i++) {
		page_map->remapped[i] = agp_bridge->scratch_page_addr;
	}

	return 0;
}

static void amd_free_page_map(amd_page_map *page_map)
{
	iounmap(page_map->remapped);
#ifdef CONFIG_X86
	change_page_attr(virt_to_page(page_map->real), 1, PAGE_KERNEL);
#endif
	ClearPageReserved(virt_to_page(page_map->real));
	free_page((unsigned long) page_map->real);
}

static void amd_free_gatt_pages(void)
{
	int i;
	amd_page_map **tables;
	amd_page_map *entry;

	tables = amd_irongate_private.gatt_pages;
	for(i = 0; i < amd_irongate_private.num_tables; i++) {
		entry = tables[i];
		if (entry != NULL) {
			if (entry->real != NULL) {
				amd_free_page_map(entry);
			}
			kfree(entry);
		}
	}
	kfree(tables);
}

static int amd_create_gatt_pages(int nr_tables)
{
	amd_page_map **tables;
	amd_page_map *entry;
	int retval = 0;
	int i;

	tables = kmalloc((nr_tables + 1) * sizeof(amd_page_map *), 
			 GFP_KERNEL);
	if (tables == NULL) {
		return -ENOMEM;
	}
	memset(tables, 0, sizeof(amd_page_map *) * (nr_tables + 1));
	for (i = 0; i < nr_tables; i++) {
		entry = kmalloc(sizeof(amd_page_map), GFP_KERNEL);
		if (entry == NULL) {
			retval = -ENOMEM;
			break;
		}
		memset(entry, 0, sizeof(amd_page_map));
		tables[i] = entry;
		retval = amd_create_page_map(entry);
		if (retval != 0) break;
	}
	amd_irongate_private.num_tables = nr_tables;
	amd_irongate_private.gatt_pages = tables;

	if (retval != 0) amd_free_gatt_pages();

	return retval;
}

/* Since we don't need contigious memory we just try
 * to get the gatt table once
 */

#define GET_PAGE_DIR_OFF(addr) (addr >> 22)
#define GET_PAGE_DIR_IDX(addr) (GET_PAGE_DIR_OFF(addr) - \
	GET_PAGE_DIR_OFF(agp_bridge->gart_bus_addr))
#define GET_GATT_OFF(addr) ((addr & 0x003ff000) >> 12) 
#define GET_GATT(addr) (amd_irongate_private.gatt_pages[\
	GET_PAGE_DIR_IDX(addr)]->remapped)

static int amd_create_gatt_table(void)
{
	aper_size_info_lvl2 *value;
	amd_page_map page_dir;
	unsigned long addr;
	int retval;
	u32 temp;
	int i;

	value = A_SIZE_LVL2(agp_bridge->current_size);
	retval = amd_create_page_map(&page_dir);
	if (retval != 0) {
		return retval;
	}

	retval = amd_create_gatt_pages(value->num_entries / 1024);
	if (retval != 0) {
		amd_free_page_map(&page_dir);
		return retval;
	}

	agp_bridge->gatt = NULL;
	agp_bridge->gatt_table = NULL;
	agp_bridge->gatt_bus_addr = virt_to_bus(page_dir.real);

	/* Get the address for the gart region.
	 * This is a bus address even on the alpha, b/c its
	 * used to program the agp master not the cpu
	 */

	pci_read_config_dword(agp_bridge->dev, AMD_APBASE, &temp);
	addr = (temp & PCI_BASE_ADDRESS_MEM_MASK);
	agp_bridge->gart_bus_addr = addr;

	/* Calculate the agp offset */
	for(i = 0; i < value->num_entries / 1024; i++, addr += 0x00400000) {
		page_dir.remapped[GET_PAGE_DIR_OFF(addr)] =
			virt_to_bus(amd_irongate_private.gatt_pages[i]->real);
		page_dir.remapped[GET_PAGE_DIR_OFF(addr)] |= 0x00000001;
	}

	return 0;
}

static int amd_free_gatt_table(void)
{
	amd_page_map page_dir;

	page_dir.real = amd_irongate_private.real;
	page_dir.remapped = amd_irongate_private.remapped;

	amd_free_gatt_pages();
	amd_free_page_map(&page_dir);
	return 0;
}

static int amd_irongate_fetch_size(void)
{
	int i;
	u32 temp;
	aper_size_info_lvl2 *values;

	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
	temp = (temp & 0x0000000e);
	values = A_SIZE_LVL2(agp_bridge->aperture_sizes);
	for (i = 0; i < agp_bridge->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;
}

static int amd_irongate_configure(void)
{
	aper_size_info_lvl2 *current_size;
	u32 temp;
	u16 enable_reg;

	current_size = A_SIZE_LVL2(agp_bridge->current_size);

	/* Get the memory mapped registers */
	pci_read_config_dword(agp_bridge->dev, AMD_MMBASE, &temp);
	temp = (temp & PCI_BASE_ADDRESS_MEM_MASK);
	amd_irongate_private.registers = (volatile u8 *) ioremap(temp, 4096);

	/* Write out the address of the gatt table */
	OUTREG32(amd_irongate_private.registers, AMD_ATTBASE,
		 agp_bridge->gatt_bus_addr);

	/* Write the Sync register */
	pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL, 0x80);
   
   	/* Set indexing mode */
   	pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL2, 0x00);

	/* Write the enable register */
	enable_reg = INREG16(amd_irongate_private.registers, AMD_GARTENABLE);
	enable_reg = (enable_reg | 0x0004);
	OUTREG16(amd_irongate_private.registers, AMD_GARTENABLE, enable_reg);

	/* Write out the size register */
	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
	temp = (((temp & ~(0x0000000e)) | current_size->size_value)
		| 0x00000001);
	pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);

	/* Flush the tlb */
	OUTREG32(amd_irongate_private.registers, AMD_TLBFLUSH, 0x00000001);

	return 0;
}

static void amd_irongate_cleanup(void)
{
	aper_size_info_lvl2 *previous_size;
	u32 temp;
	u16 enable_reg;

	previous_size = A_SIZE_LVL2(agp_bridge->previous_size);

	enable_reg = INREG16(amd_irongate_private.registers, AMD_GARTENABLE);
	enable_reg = (enable_reg & ~(0x0004));
	OUTREG16(amd_irongate_private.registers, AMD_GARTENABLE, enable_reg);

	/* Write back the previous size and disable gart translation */
	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
	temp = ((temp & ~(0x0000000f)) | previous_size->size_value);
	pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
	iounmap((void *) amd_irongate_private.registers);
}

/*
 * This routine could be implemented by taking the addresses
 * written to the GATT, and flushing them individually.  However
 * currently it just flushes the whole table.  Which is probably
 * more efficent, since agp_memory blocks can be a large number of
 * entries.
 */

static void amd_irongate_tlbflush(agp_memory * temp)
{
	OUTREG32(amd_irongate_private.registers, AMD_TLBFLUSH, 0x00000001);
}

static int amd_insert_memory(agp_memory * mem,
			     off_t pg_start, int type)
{
	int i, j, num_entries;
	unsigned long *cur_gatt;
	unsigned long addr;

	num_entries = A_SIZE_LVL2(agp_bridge->current_size)->num_entries;

	if (type != 0 || mem->type != 0) {
		return -EINVAL;
	}
	if ((pg_start + mem->page_count) > num_entries) {
		return -EINVAL;
	}

	if(agp_test_region_for_bound_intersect(pg_start, mem->page_count)) {
		return -EBUSY;
	}

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

	for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
		addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
		cur_gatt = GET_GATT(addr);
		cur_gatt[GET_GATT_OFF(addr)] = mem->memory[i];
	}
	agp_bridge->tlb_flush(mem);
	return 0;
}

static int amd_remove_memory(agp_memory * mem, off_t pg_start,
			     int type)
{
	int i;
	unsigned long *cur_gatt;
	unsigned long addr;

	if (type != 0 || mem->type != 0) {
		return -EINVAL;
	}
	for (i = pg_start; i < (mem->page_count + pg_start); i++) {
		addr = (i * PAGE_SIZE) + agp_bridge->gart_bus_addr;
		cur_gatt = GET_GATT(addr);
		cur_gatt[GET_GATT_OFF(addr)] = 
			(unsigned long) agp_bridge->scratch_page_addr;
	}

	agp_bridge->tlb_flush(mem);
	return 0;
}

static aper_size_info_lvl2 amd_irongate_sizes[7] =
{
	{2048, 524288, 0x0000000c},
	{1024, 262144, 0x0000000a},
	{512, 131072, 0x00000008},
	{256, 65536, 0x00000006},
	{128, 32768, 0x00000004},
	{64, 16384, 0x00000002},
	{32, 8192, 0x00000000}
};

static int __init amd_irongate_setup (struct pci_dev *pdev)
{
	agp_bridge->aperture_sizes = (void *) amd_irongate_sizes;
	agp_bridge->size_type = LVL2_APER_SIZE;
	agp_bridge->num_aperture_sizes = 7;
	agp_bridge->dev_private_data = (void *) &amd_irongate_private;
	agp_bridge->needs_scratch_page = FALSE;
	agp_bridge->configure = amd_irongate_configure;
	agp_bridge->fetch_size = amd_irongate_fetch_size;
	agp_bridge->cleanup = amd_irongate_cleanup;
	agp_bridge->tlb_flush = amd_irongate_tlbflush;
	agp_bridge->mask_memory = agp_generic_mask_memory;
	agp_bridge->agp_enable = agp_generic_agp_enable;
	agp_bridge->cache_flush = global_cache_flush;
	agp_bridge->create_gatt_table = amd_create_gatt_table;
	agp_bridge->free_gatt_table = amd_free_gatt_table;
	agp_bridge->insert_memory = amd_insert_memory;
	agp_bridge->remove_memory = amd_remove_memory;
	agp_bridge->alloc_by_type = agp_generic_alloc_by_type;
	agp_bridge->free_by_type = agp_generic_free_by_type;
	agp_bridge->agp_alloc_page = agp_generic_alloc_page;
	agp_bridge->agp_free_page = agp_generic_free_page;
	agp_bridge->get_reserved_map = agp_generic_get_reserved_map;
	agp_bridge->suspend = agp_generic_suspend;
	agp_bridge->resume = agp_generic_resume;
	agp_bridge->get_vm_config = agp_generic_get_vm_config;

	return 0;
	
	(void) pdev; /* unused */
}

#endif /* CONFIG_AGP_AMD */

#ifdef CONFIG_AGP_AMD_8151

/* Begin AMD-8151 support */

static u_int64_t pci_read64 (struct pci_dev *dev, int reg)
{
	union {
		u64 full;
		struct {
			u32 high;
			u32 low;
		} split;
	} tmp;
	pci_read_config_dword(dev, reg, &tmp.split.high);
	pci_read_config_dword(dev, reg+4, &tmp.split.low);
	return tmp.full;
}

static void pci_write64 (struct pci_dev *dev, int reg, u64 value)
{
	union {
		u64 full;
		struct {
			u32 high;
			u32 low;
		} split;
	} tmp;
	tmp.full = value;
	pci_write_config_dword(dev, reg, tmp.split.high);
	pci_write_config_dword(dev, reg+4, tmp.split.low);
}


static int x86_64_insert_memory(agp_memory * mem, off_t pg_start, int type)
{
	int i, j, num_entries;
	void *temp;
	long tmp;
	u32 pte;
	u64 addr;

	temp = agp_bridge->current_size;

	num_entries = A_SIZE_32(temp)->num_entries;

	num_entries -= agp_memory_reserved>>PAGE_SHIFT;

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

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

	j = pg_start;

	/* gatt table should be empty. */
	if(agp_test_region_for_bound_intersect(pg_start, mem->page_count)) {
		return -EBUSY;
	}

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

	for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
		addr = mem->memory[i];

		tmp = addr;
		BUG_ON(tmp & 0xffffff0000000ffc);
		pte = (tmp & 0x000000ff00000000) >> 28;
		pte |=(tmp & 0x00000000fffff000);
		pte |= 1<<1|1<<0;

		agp_bridge->gatt_table[j] = pte;
	}
	agp_bridge->tlb_flush(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 aper_size_info_32 x86_64_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 amd_x86_64_fetch_size(void)
{
	struct pci_dev *dev;
	int i;
	u32 temp;
	aper_size_info_32 *values;

	pci_for_each_dev(dev) {
		if (dev->bus->number==0 &&
			PCI_FUNC(dev->devfn)==3 &&
			PCI_SLOT(dev->devfn)>=24 && PCI_SLOT(dev->devfn)<=31) {

			pci_read_config_dword(dev, AMD_X86_64_GARTAPERTURECTL, &temp);
			temp = (temp & 0xe);
			values = A_SIZE_32(x86_64_aperture_sizes);

			for (i = 0; i < agp_bridge->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;
				}
			}
		}
	}
	/* erk, couldn't find an x86-64 ? */
	return 0;
}


static void inline flush_x86_64_tlb(struct pci_dev *dev)
{
	u32 tmp;

	pci_read_config_dword (dev, AMD_X86_64_GARTCACHECTL, &tmp);
	tmp |= 1<<0;
	pci_write_config_dword (dev, AMD_X86_64_GARTCACHECTL, tmp);
}


void amd_x86_64_tlbflush(agp_memory * temp)
{
	struct pci_dev *dev;

	pci_for_each_dev(dev) {
		if (dev->bus->number==0 && PCI_FUNC(dev->devfn)==3 &&
		    PCI_SLOT(dev->devfn) >=24 && PCI_SLOT(dev->devfn) <=31) {
			flush_x86_64_tlb (dev);
		}
	}
}


/*
 * In a multiprocessor x86-64 system, this function gets
 * called once for each CPU.
 */
u64 amd_x86_64_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, AMD_X86_64_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, AMD_X86_64_GARTTABLEBASE, tmp);

	/* Enable GART translation for this hammer. */
	pci_read_config_dword(hammer, AMD_X86_64_GARTAPERTURECTL, &tmp);
	tmp &= 0x3f;
	tmp |= 1<<0;
	pci_write_config_dword(hammer, AMD_X86_64_GARTAPERTURECTL, tmp);

	/* keep CPU's coherent. */
	flush_x86_64_tlb (hammer);
	
	return aper_base;
}


static 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)
{
	struct pci_dev *dev, *hammer=NULL;
	int current_size;
	int tmp, tmp2, i;
	u64 aperbar;
	unsigned long gatt_bus = __pa(page_address(agp_bridge->gatt));

	/* Configure AGP regs in each x86-64 host bridge. */
	pci_for_each_dev(dev) {
		if (dev->bus->number==0 &&
			PCI_FUNC(dev->devfn)==3 &&
			PCI_SLOT(dev->devfn)>=24 && PCI_SLOT(dev->devfn)<=31) {
			agp_bridge->gart_bus_addr = amd_x86_64_configure(dev,gatt_bus);
			hammer = dev;

			/*
			 * TODO: Cache pci_dev's of x86-64's in private struct to save us
			 * having to scan the pci list each time.
			 */
		}
	}

	if (hammer == NULL) {
		return -ENODEV;
	}

	/* Shadow x86-64 registers into 8151 registers. */

	dev = agp_bridge->dev;
	if (!dev) 
		return -ENODEV;

	current_size = amd_x86_64_fetch_size();

	pci_read_config_dword(dev, AMD_8151_APERTURESIZE, &tmp);
	tmp &= ~(0xfff);

	/* translate x86-64 size bits to 8151 size bits*/
	for (i=0 ; i<7; i++) {
		if (amd_8151_sizes[i].size == current_size)
			tmp |= (amd_8151_sizes[i].size_value) << 3;
	}
	pci_write_config_dword(dev, AMD_8151_APERTURESIZE, tmp);

	pci_read_config_dword (hammer, AMD_X86_64_GARTAPERTUREBASE, &tmp);
	aperbar = pci_read64 (dev, AMD_8151_VMAPERTURE);
	aperbar |= (tmp & 0x7fff) <<25;
	aperbar &= 0x000000ffffffffff;
	aperbar |= 1<<2;	/* This address is a 64bit ptr FIXME: Make conditional in 32bit mode */
	pci_write64 (dev, AMD_8151_VMAPERTURE, aperbar);

	pci_read_config_dword(dev, AMD_8151_AGP_CTL , &tmp);
	tmp &= ~(AMD_8151_GTLBEN | AMD_8151_APEREN);
	
	pci_read_config_dword(hammer, AMD_X86_64_GARTAPERTURECTL, &tmp2);
	if (tmp2 & AMD_X86_64_GARTEN)
		tmp |= AMD_8151_APEREN;
	// FIXME: bit 7 of AMD_8151_AGP_CTL (GTLBEN) must be copied if set.
	// But where is it set ?
	pci_write_config_dword(dev, AMD_8151_AGP_CTL, tmp);

	return 0;
}


static void amd_8151_cleanup(void)
{
	struct pci_dev *dev;
	u32 tmp;

	pci_for_each_dev(dev) {
		/* disable gart translation */
		if (dev->bus->number==0 && PCI_FUNC(dev->devfn)==3 &&
		    (PCI_SLOT(dev->devfn) >=24) && (PCI_SLOT(dev->devfn) <=31)) {

			pci_read_config_dword (dev, AMD_X86_64_GARTAPERTURECTL, &tmp);
			tmp &= ~(AMD_X86_64_GARTEN);
			pci_write_config_dword (dev, AMD_X86_64_GARTAPERTURECTL, tmp);
		}

		/* Now shadow the disable in the 8151 */
		if (dev->vendor == PCI_VENDOR_ID_AMD &&
			dev->device == PCI_DEVICE_ID_AMD_8151_0) {

			pci_read_config_dword (dev, AMD_8151_AGP_CTL, &tmp);
			tmp &= ~(AMD_8151_APEREN);	
			pci_write_config_dword (dev, AMD_8151_AGP_CTL, tmp);
		}
	}
}

static int __init amd_8151_setup (struct pci_dev *pdev)
{
	agp_bridge->aperture_sizes = (void *) amd_8151_sizes;
	agp_bridge->size_type = U32_APER_SIZE;
	agp_bridge->num_aperture_sizes = 7;
	agp_bridge->dev_private_data = NULL;
	agp_bridge->needs_scratch_page = FALSE;
	agp_bridge->configure = amd_8151_configure;
	agp_bridge->fetch_size = amd_x86_64_fetch_size;
	agp_bridge->cleanup = amd_8151_cleanup;
	agp_bridge->tlb_flush = amd_x86_64_tlbflush;
	agp_bridge->mask_memory = agp_generic_mask_memory;
	agp_bridge->agp_enable = agp_generic_agp_enable;
	agp_bridge->cache_flush = global_cache_flush;
	agp_bridge->create_gatt_table = agp_generic_create_gatt_table;
	agp_bridge->free_gatt_table = agp_generic_free_gatt_table;
	agp_bridge->insert_memory = x86_64_insert_memory;
	agp_bridge->remove_memory = agp_generic_remove_memory;
	agp_bridge->alloc_by_type = agp_generic_alloc_by_type;
	agp_bridge->free_by_type = agp_generic_free_by_type;
	agp_bridge->agp_alloc_page = agp_generic_alloc_page;
	agp_bridge->agp_free_page = agp_generic_free_page;
	agp_bridge->get_reserved_map = agp_generic_get_reserved_map;
	agp_bridge->suspend = agp_generic_suspend;
	agp_bridge->resume = agp_generic_resume;
	agp_bridge->get_vm_config = agp_generic_get_vm_config;

	return 0;
	
	(void) pdev; /* unused */
}

#endif /* CONFIG_AGP_AMD_8151 */