Merge commit 'v2.6.28-rc2' into x86/pci-ioapic-boot-irq-quirks

1 files changed, 514 insertions, 92 deletions

diff --git a/arch/x86/kernel/amd_iommu.c b/arch/x86/kernel/amd_iommu.c
index f2766d84c7a..a8fd9ebdc8e 100644
--- a/arch/x86/kernel/amd_iommu.c
+++ b/arch/x86/kernel/amd_iommu.c
@@ -23,36 +23,149 @@
 #include <linux/scatterlist.h>
 #include <linux/iommu-helper.h>
 #include <asm/proto.h>
-#include <asm/gart.h>
+#include <asm/iommu.h>
 #include <asm/amd_iommu_types.h>
 #include <asm/amd_iommu.h>
 
 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
 
-#define to_pages(addr, size) \
-	 (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
+#define EXIT_LOOP_COUNT 10000000
 
 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
 
-struct command {
+/* A list of preallocated protection domains */
+static LIST_HEAD(iommu_pd_list);
+static DEFINE_SPINLOCK(iommu_pd_list_lock);
+
+/*
+ * 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);
 
+/* 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 & IOMMU_CAP_NPCACHE;
 }
 
-static int __iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
+/****************************************************************************
+ *
+ * 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);
+	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);
@@ -63,7 +176,11 @@ static int __iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
 	return 0;
 }
 
-static int iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
+/*
+ * 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;
@@ -75,35 +192,59 @@ static int iommu_queue_command(struct amd_iommu *iommu, struct command *cmd)
 	return ret;
 }
 
+/*
+ * 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;
-	struct command cmd;
-	volatile u64 ready = 0;
-	unsigned long ready_phys = virt_to_phys(&ready);
+	int ret = 0, ready = 0;
+	unsigned status = 0;
+	struct iommu_cmd cmd;
+	unsigned long flags, i = 0;
 
 	memset(&cmd, 0, sizeof(cmd));
-	cmd.data[0] = LOW_U32(ready_phys) | CMD_COMPL_WAIT_STORE_MASK;
-	cmd.data[1] = HIGH_U32(ready_phys);
-	cmd.data[2] = 1; /* value written to 'ready' */
+	cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
 	CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
 
 	iommu->need_sync = 0;
 
-	ret = iommu_queue_command(iommu, &cmd);
+	spin_lock_irqsave(&iommu->lock, flags);
+
+	ret = __iommu_queue_command(iommu, &cmd);
 
 	if (ret)
-		return ret;
+		goto out;
+
+	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);
 
-	while (!ready)
-		cpu_relax();
+	if (unlikely((i == EXIT_LOOP_COUNT) && printk_ratelimit()))
+		printk(KERN_WARNING "AMD IOMMU: Completion wait loop failed\n");
+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 command cmd;
+	struct iommu_cmd cmd;
+	int ret;
 
 	BUG_ON(iommu == NULL);
 
@@ -111,37 +252,50 @@ static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
 	CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
 	cmd.data[0] = devid;
 
+	ret = iommu_queue_command(iommu, &cmd);
+
 	iommu->need_sync = 1;
 
-	return iommu_queue_command(iommu, &cmd);
+	return ret;
 }
 
+/*
+ * 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 command cmd;
+	struct iommu_cmd cmd;
+	int ret;
 
 	memset(&cmd, 0, sizeof(cmd));
 	address &= PAGE_MASK;
 	CMD_SET_TYPE(&cmd, CMD_INV_IOMMU_PAGES);
 	cmd.data[1] |= domid;
-	cmd.data[2] = LOW_U32(address);
-	cmd.data[3] = HIGH_U32(address);
-	if (s)
+	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)
+	if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
 		cmd.data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
 
+	ret = iommu_queue_command(iommu, &cmd);
+
 	iommu->need_sync = 1;
 
-	return 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 = to_pages(address, size);
+	unsigned pages = iommu_num_pages(address, size, PAGE_SIZE);
 
 	address &= PAGE_MASK;
 
@@ -159,6 +313,28 @@ static int iommu_flush_pages(struct amd_iommu *iommu, u16 domid,
 	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);
+}
+
+/****************************************************************************
+ *
+ * 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(struct protection_domain *dom,
 		     unsigned long bus_addr,
 		     unsigned long phys_addr,
@@ -209,6 +385,10 @@ static int iommu_map(struct protection_domain *dom,
 	return 0;
 }
 
+/*
+ * 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)
 {
@@ -223,6 +403,12 @@ static int iommu_for_unity_map(struct amd_iommu *iommu,
 	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;
@@ -239,6 +425,10 @@ static int iommu_init_unity_mappings(struct amd_iommu *iommu)
 	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)
 {
@@ -261,6 +451,9 @@ static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
 	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)
 {
@@ -278,33 +471,48 @@ static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
 	return 0;
 }
 
-static unsigned long dma_mask_to_pages(unsigned long mask)
-{
-	return (mask >> PAGE_SHIFT) +
-		(PAGE_ALIGN(mask & ~PAGE_MASK) >> PAGE_SHIFT);
-}
+/****************************************************************************
+ *
+ * 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 int pages,
+					     unsigned long align_mask,
+					     u64 dma_mask)
 {
-	unsigned long limit = dma_mask_to_pages(*dev->dma_mask);
+	unsigned long limit;
 	unsigned long address;
-	unsigned long size = dom->aperture_size >> PAGE_SHIFT;
 	unsigned long boundary_size;
 
 	boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
 			PAGE_SIZE) >> PAGE_SHIFT;
-	limit = limit < size ? limit : size;
+	limit = iommu_device_max_index(dom->aperture_size >> PAGE_SHIFT, 0,
+				       dma_mask >> PAGE_SHIFT);
 
-	if (dom->next_bit >= limit)
+	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, 0);
-	if (address == -1)
+				   0 , boundary_size, align_mask);
+	if (address == -1) {
 		address = iommu_area_alloc(dom->bitmap, limit, 0, pages,
-				0, boundary_size, 0);
+				0, boundary_size, align_mask);
+		dom->need_flush = true;
+	}
 
 	if (likely(address != -1)) {
 		dom->next_bit = address + pages;
@@ -317,6 +525,11 @@ static unsigned long dma_ops_alloc_addresses(struct device *dev,
 	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)
@@ -325,6 +538,16 @@ static void dma_ops_free_addresses(struct dma_ops_domain *dom,
 	iommu_area_free(dom->bitmap, address, pages);
 }
 
+/****************************************************************************
+ *
+ * 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;
@@ -342,6 +565,10 @@ static u16 domain_id_alloc(void)
 	return id;
 }
 
+/*
+ * 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)
@@ -351,7 +578,7 @@ static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
 	if (start_page + pages > last_page)
 		pages = last_page - start_page;
 
-	set_bit_string(dom->bitmap, start_page, pages);
+	iommu_area_reserve(dom->bitmap, start_page, pages);
 }
 
 static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
@@ -382,6 +609,10 @@ static void dma_ops_free_pagetable(struct dma_ops_domain *dma_dom)
 	free_page((unsigned long)p1);
 }
 
+/*
+ * 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)
@@ -396,6 +627,11 @@ static void dma_ops_domain_free(struct dma_ops_domain *dom)
 	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)
 {
@@ -436,14 +672,24 @@ static struct dma_ops_domain *dma_ops_domain_alloc(struct amd_iommu *iommu,
 	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 = to_pages(iommu->exclusion_start,
-				iommu->exclusion_length);
+		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);
@@ -472,6 +718,10 @@ free_dma_dom:
 	return NULL;
 }
 
+/*
+ * 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;
@@ -484,6 +734,10 @@ static struct protection_domain *domain_for_device(u16 devid)
 	return dom;
 }
 
+/*
+ * If a device is not yet associated with a domain, this function does
+ * assigns it visible for the hardware
+ */
 static void set_device_domain(struct amd_iommu *iommu,
 			      struct protection_domain *domain,
 			      u16 devid)
@@ -492,12 +746,13 @@ static void set_device_domain(struct amd_iommu *iommu,
 
 	u64 pte_root = virt_to_phys(domain->pt_root);
 
-	pte_root |= (domain->mode & 0x07) << 9;
-	pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | 2;
+	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] = pte_root;
-	amd_iommu_dev_table[devid].data[1] = pte_root >> 32;
+	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;
@@ -508,6 +763,58 @@ static void set_device_domain(struct amd_iommu *iommu,
 	iommu->need_sync = 1;
 }
 
+/*****************************************************************************
+ *
+ * The next functions belong to the dma_ops mapping/unmapping code.
+ *
+ *****************************************************************************/
+
+/*
+ * This function checks if the driver got a valid device from the caller to
+ * avoid dereferencing invalid pointers.
+ */
+static bool check_device(struct device *dev)
+{
+	if (!dev || !dev->dma_mask)
+		return false;
+
+	return true;
+}
+
+/*
+ * In this function the list of preallocated protection domains is traversed to
+ * find the domain for a specific device
+ */
+static struct dma_ops_domain *find_protection_domain(u16 devid)
+{
+	struct dma_ops_domain *entry, *ret = NULL;
+	unsigned long flags;
+
+	if (list_empty(&iommu_pd_list))
+		return NULL;
+
+	spin_lock_irqsave(&iommu_pd_list_lock, flags);
+
+	list_for_each_entry(entry, &iommu_pd_list, list) {
+		if (entry->target_dev == devid) {
+			ret = entry;
+			list_del(&ret->list);
+			break;
+		}
+	}
+
+	spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
+
+	return ret;
+}
+
+/*
+ * In the dma_ops path we only have the struct device. This function
+ * finds the corresponding IOMMU, the protection domain and the
+ * requestor id for a given device.
+ * If the device is not yet associated with a domain this is also done
+ * in this function.
+ */
 static int get_device_resources(struct device *dev,
 				struct amd_iommu **iommu,
 				struct protection_domain **domain,
@@ -517,26 +824,30 @@ static int get_device_resources(struct device *dev,
 	struct pci_dev *pcidev;
 	u16 _bdf;
 
-	BUG_ON(!dev || dev->bus != &pci_bus_type || !dev->dma_mask);
+	*iommu = NULL;
+	*domain = NULL;
+	*bdf = 0xffff;
+
+	if (dev->bus != &pci_bus_type)
+		return 0;
 
 	pcidev = to_pci_dev(dev);
-	_bdf = (pcidev->bus->number << 8) | pcidev->devfn;
+	_bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
 
-	if (_bdf >= amd_iommu_last_bdf) {
-		*iommu = NULL;
-		*domain = NULL;
-		*bdf = 0xffff;
+	/* device not translated by any IOMMU in the system? */
+	if (_bdf > amd_iommu_last_bdf)
 		return 0;
-	}
 
 	*bdf = amd_iommu_alias_table[_bdf];
 
 	*iommu = amd_iommu_rlookup_table[*bdf];
 	if (*iommu == NULL)
 		return 0;
-	dma_dom = (*iommu)->default_dom;
 	*domain = domain_for_device(*bdf);
 	if (*domain == NULL) {
+		dma_dom = find_protection_domain(*bdf);
+		if (!dma_dom)
+			dma_dom = (*iommu)->default_dom;
 		*domain = &dma_dom->domain;
 		set_device_domain(*iommu, *domain, *bdf);
 		printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
@@ -547,6 +858,10 @@ static int get_device_resources(struct device *dev,
 	return 1;
 }
 
+/*
+ * This is the generic map function. It maps one 4kb page at paddr to
+ * the given address in the DMA address space for the domain.
+ */
 static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
 				     struct dma_ops_domain *dom,
 				     unsigned long address,
@@ -578,6 +893,9 @@ static dma_addr_t dma_ops_domain_map(struct amd_iommu *iommu,
 	return (dma_addr_t)address;
 }
 
+/*
+ * The generic unmapping function for on page in the DMA address space.
+ */
 static void dma_ops_domain_unmap(struct amd_iommu *iommu,
 				 struct dma_ops_domain *dom,
 				 unsigned long address)
@@ -597,22 +915,35 @@ static void dma_ops_domain_unmap(struct amd_iommu *iommu,
 	*pte = 0ULL;
 }
 
+/*
+ * This function contains common code for mapping of a physically
+ * contiguous memory region into DMA address space. It is uses by all
+ * mapping functions provided by this IOMMU driver.
+ * Must be called with the domain lock held.
+ */
 static dma_addr_t __map_single(struct device *dev,
 			       struct amd_iommu *iommu,
 			       struct dma_ops_domain *dma_dom,
 			       phys_addr_t paddr,
 			       size_t size,
-			       int dir)
+			       int dir,
+			       bool align,
+			       u64 dma_mask)
 {
 	dma_addr_t offset = paddr & ~PAGE_MASK;
 	dma_addr_t address, start;
 	unsigned int pages;
+	unsigned long align_mask = 0;
 	int i;
 
-	pages = to_pages(paddr, size);
+	pages = iommu_num_pages(paddr, size, PAGE_SIZE);
 	paddr &= PAGE_MASK;
 
-	address = dma_ops_alloc_addresses(dev, dma_dom, pages);
+	if (align)
+		align_mask = (1UL << get_order(size)) - 1;
+
+	address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
+					  dma_mask);
 	if (unlikely(address == bad_dma_address))
 		goto out;
 
@@ -624,10 +955,20 @@ static dma_addr_t __map_single(struct device *dev,
 	}
 	address += offset;
 
+	if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
+		iommu_flush_tlb(iommu, dma_dom->domain.id);
+		dma_dom->need_flush = false;
+	} else if (unlikely(iommu_has_npcache(iommu)))
+		iommu_flush_pages(iommu, dma_dom->domain.id, address, size);
+
 out:
 	return address;
 }
 
+/*
+ * Does the reverse of the __map_single function. Must be called with
+ * the domain lock held too
+ */
 static void __unmap_single(struct amd_iommu *iommu,
 			   struct dma_ops_domain *dma_dom,
 			   dma_addr_t dma_addr,
@@ -640,7 +981,7 @@ static void __unmap_single(struct amd_iommu *iommu,
 	if ((dma_addr == 0) || (dma_addr + size > dma_dom->aperture_size))
 		return;
 
-	pages = to_pages(dma_addr, size);
+	pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
 	dma_addr &= PAGE_MASK;
 	start = dma_addr;
 
@@ -650,8 +991,14 @@ static void __unmap_single(struct amd_iommu *iommu,
 	}
 
 	dma_ops_free_addresses(dma_dom, dma_addr, pages);
+
+	if (amd_iommu_unmap_flush)
+		iommu_flush_pages(iommu, dma_dom->domain.id, dma_addr, size);
 }
 
+/*
+ * The exported map_single function for dma_ops.
+ */
 static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
 			     size_t size, int dir)
 {
@@ -660,21 +1007,26 @@ static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
 	struct protection_domain *domain;
 	u16 devid;
 	dma_addr_t addr;
+	u64 dma_mask;
+
+	if (!check_device(dev))
+		return bad_dma_address;
+
+	dma_mask = *dev->dma_mask;
 
 	get_device_resources(dev, &iommu, &domain, &devid);
 
 	if (iommu == NULL || domain == NULL)
+		/* device not handled by any AMD IOMMU */
 		return (dma_addr_t)paddr;
 
 	spin_lock_irqsave(&domain->lock, flags);
-	addr = __map_single(dev, iommu, domain->priv, paddr, size, dir);
+	addr = __map_single(dev, iommu, domain->priv, paddr, size, dir, false,
+			    dma_mask);
 	if (addr == bad_dma_address)
 		goto out;
 
-	if (iommu_has_npcache(iommu))
-		iommu_flush_pages(iommu, domain->id, addr, size);
-
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 out:
@@ -683,6 +1035,9 @@ out:
 	return addr;
 }
 
+/*
+ * The exported unmap_single function for dma_ops.
+ */
 static void unmap_single(struct device *dev, dma_addr_t dma_addr,
 			 size_t size, int dir)
 {
@@ -691,21 +1046,25 @@ static void unmap_single(struct device *dev, dma_addr_t dma_addr,
 	struct protection_domain *domain;
 	u16 devid;
 
-	if (!get_device_resources(dev, &iommu, &domain, &devid))
+	if (!check_device(dev) ||
+	    !get_device_resources(dev, &iommu, &domain, &devid))
+		/* device not handled by any AMD IOMMU */
 		return;
 
 	spin_lock_irqsave(&domain->lock, flags);
 
 	__unmap_single(iommu, domain->priv, dma_addr, size, dir);
 
-	iommu_flush_pages(iommu, domain->id, dma_addr, size);
-
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 	spin_unlock_irqrestore(&domain->lock, flags);
 }
 
+/*
+ * This is a special map_sg function which is used if we should map a
+ * device which is not handled by an AMD IOMMU in the system.
+ */
 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
 			   int nelems, int dir)
 {
@@ -720,6 +1079,10 @@ static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
 	return nelems;
 }
 
+/*
+ * The exported map_sg function for dma_ops (handles scatter-gather
+ * lists).
+ */
 static int map_sg(struct device *dev, struct scatterlist *sglist,
 		  int nelems, int dir)
 {
@@ -731,6 +1094,12 @@ static int map_sg(struct device *dev, struct scatterlist *sglist,
 	struct scatterlist *s;
 	phys_addr_t paddr;
 	int mapped_elems = 0;
+	u64 dma_mask;
+
+	if (!check_device(dev))
+		return 0;
+
+	dma_mask = *dev->dma_mask;
 
 	get_device_resources(dev, &iommu, &domain, &devid);
 
@@ -743,19 +1112,17 @@ static int map_sg(struct device *dev, struct scatterlist *sglist,
 		paddr = sg_phys(s);
 
 		s->dma_address = __map_single(dev, iommu, domain->priv,
-					      paddr, s->length, dir);
+					      paddr, s->length, dir, false,
+					      dma_mask);
 
 		if (s->dma_address) {
 			s->dma_length = s->length;
 			mapped_elems++;
 		} else
 			goto unmap;
-		if (iommu_has_npcache(iommu))
-			iommu_flush_pages(iommu, domain->id, s->dma_address,
-					  s->dma_length);
 	}
 
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 out:
@@ -775,6 +1142,10 @@ unmap:
 	goto out;
 }
 
+/*
+ * The exported map_sg function for dma_ops (handles scatter-gather
+ * lists).
+ */
 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
 		     int nelems, int dir)
 {
@@ -785,7 +1156,8 @@ static void unmap_sg(struct device *dev, struct scatterlist *sglist,
 	u16 devid;
 	int i;
 
-	if (!get_device_resources(dev, &iommu, &domain, &devid))
+	if (!check_device(dev) ||
+	    !get_device_resources(dev, &iommu, &domain, &devid))
 		return;
 
 	spin_lock_irqsave(&domain->lock, flags);
@@ -793,17 +1165,18 @@ static void unmap_sg(struct device *dev, struct scatterlist *sglist,
 	for_each_sg(sglist, s, nelems, i) {
 		__unmap_single(iommu, domain->priv, s->dma_address,
 			       s->dma_length, dir);
-		iommu_flush_pages(iommu, domain->id, s->dma_address,
-				  s->dma_length);
 		s->dma_address = s->dma_length = 0;
 	}
 
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 	spin_unlock_irqrestore(&domain->lock, flags);
 }
 
+/*
+ * The exported alloc_coherent function for dma_ops.
+ */
 static void *alloc_coherent(struct device *dev, size_t size,
 			    dma_addr_t *dma_addr, gfp_t flag)
 {
@@ -813,25 +1186,33 @@ static void *alloc_coherent(struct device *dev, size_t size,
 	struct protection_domain *domain;
 	u16 devid;
 	phys_addr_t paddr;
+	u64 dma_mask = dev->coherent_dma_mask;
+
+	if (!check_device(dev))
+		return NULL;
+
+	if (!get_device_resources(dev, &iommu, &domain, &devid))
+		flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
 
+	flag |= __GFP_ZERO;
 	virt_addr = (void *)__get_free_pages(flag, get_order(size));
 	if (!virt_addr)
 		return 0;
 
-	memset(virt_addr, 0, size);
 	paddr = virt_to_phys(virt_addr);
 
-	get_device_resources(dev, &iommu, &domain, &devid);
-
 	if (!iommu || !domain) {
 		*dma_addr = (dma_addr_t)paddr;
 		return virt_addr;
 	}
 
+	if (!dma_mask)
+		dma_mask = *dev->dma_mask;
+
 	spin_lock_irqsave(&domain->lock, flags);
 
 	*dma_addr = __map_single(dev, iommu, domain->priv, paddr,
-				 size, DMA_BIDIRECTIONAL);
+				 size, DMA_BIDIRECTIONAL, true, dma_mask);
 
 	if (*dma_addr == bad_dma_address) {
 		free_pages((unsigned long)virt_addr, get_order(size));
@@ -839,10 +1220,7 @@ static void *alloc_coherent(struct device *dev, size_t size,
 		goto out;
 	}
 
-	if (iommu_has_npcache(iommu))
-		iommu_flush_pages(iommu, domain->id, *dma_addr, size);
-
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 out:
@@ -851,6 +1229,9 @@ out:
 	return virt_addr;
 }
 
+/*
+ * The exported free_coherent function for dma_ops.
+ */
 static void free_coherent(struct device *dev, size_t size,
 			  void *virt_addr, dma_addr_t dma_addr)
 {
@@ -859,6 +1240,9 @@ static void free_coherent(struct device *dev, size_t size,
 	struct protection_domain *domain;
 	u16 devid;
 
+	if (!check_device(dev))
+		return;
+
 	get_device_resources(dev, &iommu, &domain, &devid);
 
 	if (!iommu || !domain)
@@ -867,9 +1251,8 @@ static void free_coherent(struct device *dev, size_t size,
 	spin_lock_irqsave(&domain->lock, flags);
 
 	__unmap_single(iommu, domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
-	iommu_flush_pages(iommu, domain->id, dma_addr, size);
 
-	if (iommu->need_sync)
+	if (unlikely(iommu->need_sync))
 		iommu_completion_wait(iommu);
 
 	spin_unlock_irqrestore(&domain->lock, flags);
@@ -879,6 +1262,32 @@ free_mem:
 }
 
 /*
+ * This function is called by the DMA layer to find out if we can handle a
+ * particular device. It is part of the dma_ops.
+ */
+static int amd_iommu_dma_supported(struct device *dev, u64 mask)
+{
+	u16 bdf;
+	struct pci_dev *pcidev;
+
+	/* No device or no PCI device */
+	if (!dev || dev->bus != &pci_bus_type)
+		return 0;
+
+	pcidev = to_pci_dev(dev);
+
+	bdf = calc_devid(pcidev->bus->number, pcidev->devfn);
+
+	/* Out of our scope? */
+	if (bdf > amd_iommu_last_bdf)
+		return 0;
+
+	return 1;
+}
+
+/*
+ * The function for pre-allocating protection domains.
+ *
  * If the driver core informs the DMA layer if a driver grabs a device
  * we don't need to preallocate the protection domains anymore.
  * For now we have to.
@@ -893,7 +1302,7 @@ void prealloc_protection_domains(void)
 
 	while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
 		devid = (dev->bus->number << 8) | dev->devfn;
-		if (devid >= amd_iommu_last_bdf)
+		if (devid > amd_iommu_last_bdf)
 			continue;
 		devid = amd_iommu_alias_table[devid];
 		if (domain_for_device(devid))
@@ -905,10 +1314,9 @@ void prealloc_protection_domains(void)
 		if (!dma_dom)
 			continue;
 		init_unity_mappings_for_device(dma_dom, devid);
-		set_device_domain(iommu, &dma_dom->domain, devid);
-		printk(KERN_INFO "AMD IOMMU: Allocated domain %d for device ",
-		       dma_dom->domain.id);
-		print_devid(devid, 1);
+		dma_dom->target_dev = devid;
+
+		list_add_tail(&dma_dom->list, &iommu_pd_list);
 	}
 }
 
@@ -919,14 +1327,23 @@ static struct dma_mapping_ops amd_iommu_dma_ops = {
 	.unmap_single = unmap_single,
 	.map_sg = map_sg,
 	.unmap_sg = unmap_sg,
+	.dma_supported = amd_iommu_dma_supported,
 };
 
+/*
+ * The function which clues the AMD IOMMU driver into dma_ops.
+ */
 int __init amd_iommu_init_dma_ops(void)
 {
 	struct amd_iommu *iommu;
 	int order = amd_iommu_aperture_order;
 	int ret;
 
+	/*
+	 * first allocate a default protection domain for every IOMMU we
+	 * found in the system. Devices not assigned to any other
+	 * protection domain will be assigned to the default one.
+	 */
 	list_for_each_entry(iommu, &amd_iommu_list, list) {
 		iommu->default_dom = dma_ops_domain_alloc(iommu, order);
 		if (iommu->default_dom == NULL)
@@ -936,6 +1353,10 @@ int __init amd_iommu_init_dma_ops(void)
 			goto free_domains;
 	}
 
+	/*
+	 * If device isolation is enabled, pre-allocate the protection
+	 * domains for each device.
+	 */
 	if (amd_iommu_isolate)
 		prealloc_protection_domains();
 
@@ -947,6 +1368,7 @@ int __init amd_iommu_init_dma_ops(void)
 	gart_iommu_aperture = 0;
 #endif
 
+	/* Make the driver finally visible to the drivers */
 	dma_ops = &amd_iommu_dma_ops;
 
 	return 0;