amd_iommu.c

/*
 * Copyright (C) 2007-2008 Advanced Micro Devices, Inc.
 * Author: Joerg Roedel <joerg.roedel@amd.com>
 *         Leo Duran <leo.duran@amd.com>
 *
 * This program is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 as published
 * by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
 */

#include <linux/pci.h>
#include <linux/gfp.h>
#include <linux/bitops.h>
#include <linux/debugfs.h>
#include <linux/scatterlist.h>
#include <linux/iommu-helper.h>
#ifdef CONFIG_IOMMU_API
#include <linux/iommu.h>
#endif
#include <asm/proto.h>
#include <asm/iommu.h>
#include <asm/gart.h>
#include <asm/amd_iommu_types.h>
#include <asm/amd_iommu.h>

#define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))

#define EXIT_LOOP_COUNT 10000000

static DEFINE_RWLOCK(amd_iommu_devtable_lock);

/* A list of preallocated protection domains */
static LIST_HEAD(iommu_pd_list);
static DEFINE_SPINLOCK(iommu_pd_list_lock);

#ifdef CONFIG_IOMMU_API
static struct iommu_ops amd_iommu_ops;
#endif

/*
 * general struct to manage commands send to an IOMMU
 */
struct iommu_cmd {
	u32 data[4];
};

static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
			     struct unity_map_entry *e);
static struct dma_ops_domain *find_protection_domain(u16 devid);


#ifdef CONFIG_AMD_IOMMU_STATS

/*
 * Initialization code for statistics collection
 */

DECLARE_STATS_COUNTER(compl_wait);

static struct dentry *stats_dir;
static struct dentry *de_isolate;
static struct dentry *de_fflush;

static void amd_iommu_stats_add(struct __iommu_counter *cnt)
{
	if (stats_dir == NULL)
		return;

	cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
				       &cnt->value);
}

static void amd_iommu_stats_init(void)
{
	stats_dir = debugfs_create_dir("amd-iommu", NULL);
	if (stats_dir == NULL)
		return;

	de_isolate = debugfs_create_bool("isolation", 0444, stats_dir,
					 (u32 *)&amd_iommu_isolate);

	de_fflush  = debugfs_create_bool("fullflush", 0444, stats_dir,
					 (u32 *)&amd_iommu_unmap_flush);

	amd_iommu_stats_add(&compl_wait);
}

#endif

/* returns !0 if the IOMMU is caching non-present entries in its TLB */
static int iommu_has_npcache(struct amd_iommu *iommu)
{
	return iommu->cap & (1UL << IOMMU_CAP_NPCACHE);
}

/****************************************************************************
 *
 * Interrupt handling functions
 *
 ****************************************************************************/

static void iommu_print_event(void *__evt)
{
	u32 *event = __evt;
	int type  = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
	int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
	int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
	int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
	u64 address = (u64)(((u64)event[3]) << 32) | event[2];

	printk(KERN_ERR "AMD IOMMU: Event logged [");

	switch (type) {
	case EVENT_TYPE_ILL_DEV:
		printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
		       "address=0x%016llx flags=0x%04x]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       address, flags);
		break;
	case EVENT_TYPE_IO_FAULT:
		printk("IO_PAGE_FAULT device=%02x:%02x.%x "
		       "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       domid, address, flags);
		break;
	case EVENT_TYPE_DEV_TAB_ERR:
		printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
		       "address=0x%016llx flags=0x%04x]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       address, flags);
		break;
	case EVENT_TYPE_PAGE_TAB_ERR:
		printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
		       "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       domid, address, flags);
		break;
	case EVENT_TYPE_ILL_CMD:
		printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
		break;
	case EVENT_TYPE_CMD_HARD_ERR:
		printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
		       "flags=0x%04x]\n", address, flags);
		break;
	case EVENT_TYPE_IOTLB_INV_TO:
		printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
		       "address=0x%016llx]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       address);
		break;
	case EVENT_TYPE_INV_DEV_REQ:
		printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
		       "address=0x%016llx flags=0x%04x]\n",
		       PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
		       address, flags);
		break;
	default:
		printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
	}
}

static void iommu_poll_events(struct amd_iommu *iommu)
{
	u32 head, tail;
	unsigned long flags;

	spin_lock_irqsave(&iommu->lock, flags);

	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);

	while (head != tail) {
		iommu_print_event(iommu->evt_buf + head);
		head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
	}

	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);

	spin_unlock_irqrestore(&iommu->lock, flags);
}

irqreturn_t amd_iommu_int_handler(int irq, void *data)
{
	struct amd_iommu *iommu;

	list_for_each_entry(iommu, &amd_iommu_list, list)
		iommu_poll_events(iommu);

	return IRQ_HANDLED;
}

/****************************************************************************
 *
 * IOMMU command queuing functions
 *
 ****************************************************************************/

/*
 * Writes the command to the IOMMUs command buffer and informs the
 * hardware about the new command. Must be called with iommu->lock held.
 */
static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
{
	u32 tail, head;
	u8 *target;

	tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
	target = iommu->cmd_buf + tail;
	memcpy_toio(target, cmd, sizeof(*cmd));
	tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
	head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
	if (tail == head)
		return -ENOMEM;
	writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);

	return 0;
}

/*
 * General queuing function for commands. Takes iommu->lock and calls
 * __iommu_queue_command().
 */
static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
{
	unsigned long flags;
	int ret;

	spin_lock_irqsave(&iommu->lock, flags);
	ret = __iommu_queue_command(iommu, cmd);
	if (!ret)
		iommu->need_sync = true;
	spin_unlock_irqrestore(&iommu->lock, flags);

	return ret;
}

/*
 * This function waits until an IOMMU has completed a completion
 * wait command
 */
static void __iommu_wait_for_completion(struct amd_iommu *iommu)
{
	int ready = 0;
	unsigned status = 0;
	unsigned long i = 0;

	INC_STATS_COUNTER(compl_wait);

	while (!ready && (i < EXIT_LOOP_COUNT)) {
		++i;
		/* wait for the bit to become one */
		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
		ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
	}

	/* set bit back to zero */
	status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
	writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);

	if (unlikely(i == EXIT_LOOP_COUNT))
		panic("AMD IOMMU: Completion wait loop failed\n");
}

/*
 * This function queues a completion wait command into the command
 * buffer of an IOMMU
 */
static int __iommu_completion_wait(struct amd_iommu *iommu)
{
	struct iommu_cmd cmd;

	 memset(&cmd, 0, sizeof(cmd));
	 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
	 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);

	 return __iommu_queue_command(iommu, &cmd);
}

/*
 * This function is called whenever we need to ensure that the IOMMU has
 * completed execution of all commands we sent. It sends a
 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
 * us about that by writing a value to a physical address we pass with
 * the command.
 */
static int iommu_completion_wait(struct amd_iommu *iommu)
{
	int ret = 0;
	unsigned long flags;

	spin_lock_irqsave(&iommu->lock, flags);

	if (!iommu->need_sync)
		goto out;

	ret = __iommu_completion_wait(iommu);

	iommu->need_sync = false;

	if (ret)
		goto out;

	__iommu_wait_for_completion(iommu);

out:
	spin_unlock_irqrestore(&iommu->lock, flags);

	return 0;
}

/*
 * Command send function for invalidating a device table entry
 */
static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
{
	struct iommu_cmd cmd;
	int ret;

	BUG_ON(iommu == NULL);

	memset(&cmd, 0, sizeof(cmd));
	CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
	cmd.data[0] = devid;

	ret = iommu_queue_command(iommu, &cmd);

	return ret;
}

static void __iommu_build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
					  u16 domid, int pde, int s)
{
	memset(cmd, 0, sizeof(*cmd));
	address &= PAGE_MASK;
	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
	cmd->data[1] |= domid;
	cmd->data[2] = lower_32_bits(address);
	cmd->data[3] = upper_32_bits(address);
	if (s) /* size bit - we flush more than one 4kb page */
		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
	if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
}

/*
 * Generic command send function for invalidaing TLB entries
 */
static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
		u64 address, u16 domid, int pde, int s)
{
	struct iommu_cmd cmd;
	int ret;

	__iommu_build_inv_iommu_pages(&cmd, address, domid, pde, s);

	ret = iommu_queue_command(iommu, &cmd);

	return ret;
}

/*
 * TLB invalidation function which is called from the mapping functions.
 * It invalidates a single PTE if the range to flush is within a single
 * page. Otherwise it flushes the whole TLB of the IOMMU.
 */
static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
		u64 address, size_t size)
{
	int s = 0;
	unsigned pages = iommu_num_pages(address, size, PAGE_SIZE);

	address &= PAGE_MASK;

	if (pages > 1) {
		/*
		 * If we have to flush more than one page, flush all
		 * TLB entries for this domain
		 */
		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
		s = 1;
	}

	iommu_queue_inv_iommu_pages(iommu, address, domid, 0, s);

	return 0;
}

/* Flush the whole IO/TLB for a given protection domain */
static void iommu_flush_tlb(struct amd_iommu *iommu, u16 domid)
{
	u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;

	iommu_queue_inv_iommu_pages(iommu, address, domid, 0, 1);
}

#ifdef CONFIG_IOMMU_API
/*
 * This function is used to flush the IO/TLB for a given protection domain
 * on every IOMMU in the system
 */
static void iommu_flush_domain(u16 domid)
{
	unsigned long flags;
	struct amd_iommu *iommu;
	struct iommu_cmd cmd;

	__iommu_build_inv_iommu_pages(&cmd, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
				      domid, 1, 1);

	list_for_each_entry(iommu, &amd_iommu_list, list) {
		spin_lock_irqsave(&iommu->lock, flags);
		__iommu_queue_command(iommu, &cmd);
		__iommu_completion_wait(iommu);
		__iommu_wait_for_completion(iommu);
		spin_unlock_irqrestore(&iommu->lock, flags);
	}
}
#endif

/****************************************************************************
 *
 * The functions below are used the create the page table mappings for
 * unity mapped regions.
 *
 ****************************************************************************/

/*
 * Generic mapping functions. It maps a physical address into a DMA
 * address space. It allocates the page table pages if necessary.
 * In the future it can be extended to a generic mapping function
 * supporting all features of AMD IOMMU page tables like level skipping
 * and full 64 bit address spaces.
 */
static int iommu_map_page(struct protection_domain *dom,
			  unsigned long bus_addr,
			  unsigned long phys_addr,
			  int prot)
{
	u64 __pte, *pte, *page;

	bus_addr  = PAGE_ALIGN(bus_addr);
	phys_addr = PAGE_ALIGN(phys_addr);

	/* only support 512GB address spaces for now */
	if (bus_addr > IOMMU_MAP_SIZE_L3 || !(prot & IOMMU_PROT_MASK))
		return -EINVAL;

	pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];

	if (!IOMMU_PTE_PRESENT(*pte)) {
		page = (u64 *)get_zeroed_page(GFP_KERNEL);
		if (!page)
			return -ENOMEM;
		*pte = IOMMU_L2_PDE(virt_to_phys(page));
	}

	pte = IOMMU_PTE_PAGE(*pte);
	pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];

	if (!IOMMU_PTE_PRESENT(*pte)) {
		page = (u64 *)get_zeroed_page(GFP_KERNEL);
		if (!page)
			return -ENOMEM;
		*pte = IOMMU_L1_PDE(virt_to_phys(page));
	}

	pte = IOMMU_PTE_PAGE(*pte);
	pte = &pte[IOMMU_PTE_L0_INDEX(bus_addr)];

	if (IOMMU_PTE_PRESENT(*pte))
		return -EBUSY;

	__pte = phys_addr | IOMMU_PTE_P;
	if (prot & IOMMU_PROT_IR)
		__pte |= IOMMU_PTE_IR;
	if (prot & IOMMU_PROT_IW)
		__pte |= IOMMU_PTE_IW;

	*pte = __pte;

	return 0;
}

#ifdef CONFIG_IOMMU_API
static void iommu_unmap_page(struct protection_domain *dom,
			     unsigned long bus_addr)
{
	u64 *pte;

	pte = &dom->pt_root[IOMMU_PTE_L2_INDEX(bus_addr)];

	if (!IOMMU_PTE_PRESENT(*pte))
		return;

	pte = IOMMU_PTE_PAGE(*pte);
	pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];

	if (!IOMMU_PTE_PRESENT(*pte))
		return;

	pte = IOMMU_PTE_PAGE(*pte);
	pte = &pte[IOMMU_PTE_L1_INDEX(bus_addr)];

	*pte = 0;
}
#endif

/*
 * This function checks if a specific unity mapping entry is needed for
 * this specific IOMMU.
 */
static int iommu_for_unity_map(struct amd_iommu *iommu,
			       struct unity_map_entry *entry)
{
	u16 bdf, i;

	for (i = entry->devid_start; i <= entry->devid_end; ++i) {
		bdf = amd_iommu_alias_table[i];
		if (amd_iommu_rlookup_table[bdf] == iommu)
			return 1;
	}

	return 0;
}

/*
 * Init the unity mappings for a specific IOMMU in the system
 *
 * Basically iterates over all unity mapping entries and applies them to
 * the default domain DMA of that IOMMU if necessary.
 */
static int iommu_init_unity_mappings(struct amd_iommu *iommu)
{
	struct unity_map_entry *entry;
	int ret;

	list_for_each_entry(entry, &amd_iommu_unity_map, list) {
		if (!iommu_for_unity_map(iommu, entry))
			continue;
		ret = dma_ops_unity_map(iommu->default_dom, entry);
		if (ret)
			return ret;
	}

	return 0;
}

/*
 * This function actually applies the mapping to the page table of the
 * dma_ops domain.
 */
static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
			     struct unity_map_entry *e)
{
	u64 addr;
	int ret;

	for (addr = e->address_start; addr < e->address_end;
	     addr += PAGE_SIZE) {
		ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot);
		if (ret)
			return ret;
		/*
		 * if unity mapping is in aperture range mark the page
		 * as allocated in the aperture
		 */
		if (addr < dma_dom->aperture_size)
			__set_bit(addr >> PAGE_SHIFT, dma_dom->bitmap);
	}

	return 0;
}

/*
 * Inits the unity mappings required for a specific device
 */
static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
					  u16 devid)
{
	struct unity_map_entry *e;
	int ret;

	list_for_each_entry(e, &amd_iommu_unity_map, list) {
		if (!(devid >= e->devid_start && devid <= e->devid_end))
			continue;
		ret = dma_ops_unity_map(dma_dom, e);
		if (ret)
			return ret;
	}

	return 0;
}

/****************************************************************************
 *
 * The next functions belong to the address allocator for the dma_ops
 * interface functions. They work like the allocators in the other IOMMU
 * drivers. Its basically a bitmap which marks the allocated pages in
 * the aperture. Maybe it could be enhanced in the future to a more
 * efficient allocator.
 *
 ****************************************************************************/

/*
 * The address allocator core function.
 *
 * called with domain->lock held
 */
static unsigned long dma_ops_alloc_addresses(struct device *dev,
					     struct dma_ops_domain *dom,
					     unsigned int pages,
					     unsigned long align_mask,
					     u64 dma_mask)
{
	unsigned long limit;
	unsigned long address;
	unsigned long boundary_size;

	boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
			PAGE_SIZE) >> PAGE_SHIFT;
	limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0,
				       dma_mask >> PAGE_SHIFT);

	if (dom->next_bit >= limit) {
		dom->next_bit = 0;
		dom->need_flush = true;
	}

	address = iommu_area_alloc(dom->bitmap, limit, dom->next_bit, pages,
				   0 , boundary_size, align_mask);
	if (address == -1) {
		address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
				0, boundary_size, align_mask);
		dom->need_flush = true;
	}

	if (likely(address != -1)) {
		dom->next_bit = address + pages;
		address <<= PAGE_SHIFT;
	} else
		address = bad_dma_address;

	WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);

	return address;
}

/*
 * The address free function.
 *
 * called with domain->lock held
 */
static void dma_ops_free_addresses(struct dma_ops_domain *dom,
				   unsigned long address,
				   unsigned int pages)
{
	address >>= PAGE_SHIFT;
	iommu_area_free(dom->bitmap, address, pages);

	if (address >= dom->next_bit)
		dom->need_flush = true;
}

/****************************************************************************
 *
 * The next functions belong to the domain allocation. A domain is
 * allocated for every IOMMU as the default domain. If device isolation
 * is enabled, every device get its own domain. The most important thing
 * about domains is the page table mapping the DMA address space they
 * contain.
 *
 ****************************************************************************/

static u16 domain_id_alloc(void)
{
	unsigned long flags;
	int id;

	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
	BUG_ON(id == 0);
	if (id > 0 && id < MAX_DOMAIN_ID)
		__set_bit(id, amd_iommu_pd_alloc_bitmap);
	else
		id = 0;
	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);

	return id;
}

#ifdef CONFIG_IOMMU_API
static void domain_id_free(int id)
{
	unsigned long flags;

	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
	if (id > 0 && id < MAX_DOMAIN_ID)
		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
}
#endif

/*
 * Used to reserve address ranges in the aperture (e.g. for exclusion
 * ranges.
 */
static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
				      unsigned long start_page,
				      unsigned int pages)
{
	unsigned int last_page = dom->aperture_size >> PAGE_SHIFT;

	if (start_page + pages > last_page)
		pages = last_page - start_page;

	iommu_area_reserve(dom->bitmap, start_page, pages);
}

static void free_pagetable(struct protection_domain *domain)
{
	int i, j;
	u64 *p1, *p2, *p3;

	p1 = domain->pt_root;

	if (!p1)
		return;

	for (i = 0; i < 512; ++i) {
		if (!IOMMU_PTE_PRESENT(p1[i]))
			continue;

		p2 = IOMMU_PTE_PAGE(p1[i]);
		for (j = 0; j < 512; ++j) {
			if (!IOMMU_PTE_PRESENT(p2[j]))
				continue;
			p3 = IOMMU_PTE_PAGE(p2[j]);
			free_page((unsigned long)p3);
		}

		free_page((unsigned long)p2);
	}

	free_page((unsigned long)p1);

	domain->pt_root = NULL;
}

/*
 * Free a domain, only used if something went wrong in the
 * allocation path and we need to free an already allocated page table
 */
static void dma_ops_domain_free(struct dma_ops_domain *dom)
{
	if (!dom)
		return;

	free_pagetable(&dom->domain);

	kfree(dom->pte_pages);

	kfree(dom->bitmap);

	kfree(dom);
}

/*
 * Allocates a new protection domain usable for the dma_ops functions.
 * It also intializes the page table and the address allocator data
 * structures required for the dma_ops interface
 */
static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
						   unsigned order)
{
	struct dma_ops_domain *dma_dom;
	unsigned i, num_pte_pages;
	u64 *l2_pde;
	u64 address;

	/*
	 * Currently the DMA aperture must be between 32 MB and 1GB in size
	 */
	if ((order < 25) || (order > 30))
		return NULL;

	dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
	if (!dma_dom)
		return NULL;

	spin_lock_init(&dma_dom->domain.lock);

	dma_dom->domain.id = domain_id_alloc();
	if (dma_dom->domain.id == 0)
		goto free_dma_dom;
	dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
	dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
	dma_dom->domain.flags = PD_DMA_OPS_MASK;
	dma_dom->domain.priv = dma_dom;
	if (!dma_dom->domain.pt_root)
		goto free_dma_dom;
	dma_dom->aperture_size = (1ULL << order);
	dma_dom->bitmap = kzalloc(dma_dom->aperture_size / (PAGE_SIZE * 8),
				  GFP_KERNEL);
	if (!dma_dom->bitmap)
		goto free_dma_dom;
	/*
	 * mark the first page as allocated so we never return 0 as
	 * a valid dma-address. So we can use 0 as error value
	 */
	dma_dom->bitmap[0] = 1;
	dma_dom->next_bit = 0;

	dma_dom->need_flush = false;
	dma_dom->target_dev = 0xffff;

	/* Intialize the exclusion range if necessary */
	if (iommu->exclusion_start &&
	    iommu->exclusion_start < dma_dom->aperture_size) {
		unsigned long startpage = iommu->exclusion_start >> PAGE_SHIFT;
		int pages = iommu_num_pages(iommu->exclusion_start,
					    iommu->exclusion_length,
					    PAGE_SIZE);
		dma_ops_reserve_addresses(dma_dom, startpage, pages);
	}

	/*
	 * At the last step, build the page tables so we don't need to
	 * allocate page table pages in the dma_ops mapping/unmapping
	 * path.
	 */
	num_pte_pages = dma_dom->aperture_size / (PAGE_SIZE * 512);
	dma_dom->pte_pages = kzalloc(num_pte_pages * sizeof(void *),
			GFP_KERNEL);
	if (!dma_dom->pte_pages)
		goto free_dma_dom;

	l2_pde = (u64 *)get_zeroed_page(GFP_KERNEL);
	if (l2_pde == NULL)
		goto free_dma_dom;

	dma_dom->domain.pt_root[0] = IOMMU_L2_PDE(virt_to_phys(l2_pde));

	for (i = 0; i < num_pte_pages; ++i) {
		dma_dom->pte_pages[i] = (u64 *)get_zeroed_page(GFP_KERNEL);
		if (!dma_dom->pte_pages[i])
			goto free_dma_dom;
		address = virt_to_phys(dma_dom->pte_pages[i]);
		l2_pde[i] = IOMMU_L1_PDE(address);
	}

	return dma_dom;

free_dma_dom:
	dma_ops_domain_free(dma_dom);

	return NULL;
}

/*
 * little helper function to check whether a given protection domain is a
 * dma_ops domain
 */
static bool dma_ops_domain(struct protection_domain *domain)
{
	return domain->flags & PD_DMA_OPS_MASK;
}

/*
 * Find out the protection domain structure for a given PCI device. This
 * will give us the pointer to the page table root for example.
 */
static struct protection_domain *domain_for_device(u16 devid)
{
	struct protection_domain *dom;
	unsigned long flags;

	read_lock_irqsave(&amd_iommu_devtable_lock, flags);
	dom = amd_iommu_pd_table[devid];
	read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);

	return dom;
}

/*
 * If a device is not yet associated with a domain, this function does
 * assigns it visible for the hardware
 */
static void attach_device(struct amd_iommu *iommu,
			  struct protection_domain *domain,
			  u16 devid)
{
	unsigned long flags;
	u64 pte_root = virt_to_phys(domain->pt_root);

	domain->dev_cnt += 1;

	pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
		    << DEV_ENTRY_MODE_SHIFT;
	pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;

	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
	amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
	amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
	amd_iommu_dev_table[devid].data[2] = domain->id;

	amd_iommu_pd_table[devid] = domain;
	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);

	iommu_queue_inv_dev_entry(iommu, devid);
}

/*
 * Removes a device from a protection domain (unlocked)
 */
static void __detach_device(struct protection_domain *domain, u16 devid)
{

	/* lock domain */
	spin_lock(&domain->lock);

	/* remove domain from the lookup table */
	amd_iommu_pd_table[devid] = NULL;

	/* remove entry from the device table seen by the hardware */
	amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
	amd_iommu_dev_table[devid].data[1] = 0;
	amd_iommu_dev_table[devid].data[2] = 0;

	/* decrease reference counter */
	domain->dev_cnt -= 1;

	/* ready */
	spin_unlock(&domain->lock);
}

/*
 * Removes a device from a protection domain (with devtable_lock held)
 */
static void detach_device(struct protection_domain *domain, u16 devid)
{
	unsigned long flags;

	/* lock device table */
	write_lock_irqsave(&amd_iommu_devtable_lock, flags);
	__detach_device(domain, devid);
	write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
}

static int device_change_notifier(struct notifier_block *nb,
				  unsigned long action, void *data)
{
	struct device *dev = data;
	struct pci_dev *pdev = to_pci_dev(dev);
	u16 devid = calc_devid(pdev->bus->number, pdev->devfn);
	struct protection_domain *domain;
	struct dma_ops_domain *dma_domain;
	struct amd_iommu *iommu;
	int order = amd_iommu_aperture_order;
	unsigned long flags;

	if (devid > amd_iommu_last_bdf)
		goto out;

	devid = amd_iommu_alias_table[devid];

	iommu = amd_iommu_rlookup_table[devid];
	if (iommu == NULL)
		goto out;

	domain = domain_for_device(devid);

	if (domain && !dma_ops_domain(domain))
		WARN_ONCE(1, "AMD IOMMU WARNING: device %s already bound "
			  "to a non-dma-ops domain\n", dev_name(dev));

	switch (action) {
	case BUS_NOTIFY_BOUND_DRIVER:
		if (domain)
			goto out;
		dma_domain = find_protection_domain(devid);
		if (!dma_domain)
			dma_domain = iommu->default_dom;
		attach_device(iommu, &dma_domain->domain, devid);
		printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
		       "device %s\n", dma_domain->domain.id, dev_name(dev));
		break;
	case BUS_NOTIFY_UNBIND_DRIVER:
		if (!domain)
			goto out;
		detach_device(domain, devid);
		break;