include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit...
[safe/jmp/linux-2.6] / arch / x86 / kernel / amd_iommu.c
index 17e83ec..f3dadb5 100644 (file)
@@ -18,8 +18,8 @@
  */
 
 #include <linux/pci.h>
-#include <linux/gfp.h>
-#include <linux/bitops.h>
+#include <linux/bitmap.h>
+#include <linux/slab.h>
 #include <linux/debugfs.h>
 #include <linux/scatterlist.h>
 #include <linux/dma-mapping.h>
@@ -57,17 +57,7 @@ struct iommu_cmd {
        u32 data[4];
 };
 
-static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
-                            struct unity_map_entry *e);
-static u64 *alloc_pte(struct protection_domain *domain,
-                     unsigned long address, int end_lvl,
-                     u64 **pte_page, gfp_t gfp);
-static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
-                                     unsigned long start_page,
-                                     unsigned int pages);
 static void reset_iommu_command_buffer(struct amd_iommu *iommu);
-static u64 *fetch_pte(struct protection_domain *domain,
-                     unsigned long address, int map_size);
 static void update_domain(struct protection_domain *domain);
 
 /****************************************************************************
@@ -83,6 +73,11 @@ static inline u16 get_device_id(struct device *dev)
        return calc_devid(pdev->bus->number, pdev->devfn);
 }
 
+static struct iommu_dev_data *get_dev_data(struct device *dev)
+{
+       return dev->archdata.iommu;
+}
+
 /*
  * In this function the list of preallocated protection domains is traversed to
  * find the domain for a specific device
@@ -138,6 +133,76 @@ static bool check_device(struct device *dev)
        return true;
 }
 
+static int iommu_init_device(struct device *dev)
+{
+       struct iommu_dev_data *dev_data;
+       struct pci_dev *pdev;
+       u16 devid, alias;
+
+       if (dev->archdata.iommu)
+               return 0;
+
+       dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
+       if (!dev_data)
+               return -ENOMEM;
+
+       dev_data->dev = dev;
+
+       devid = get_device_id(dev);
+       alias = amd_iommu_alias_table[devid];
+       pdev = pci_get_bus_and_slot(PCI_BUS(alias), alias & 0xff);
+       if (pdev)
+               dev_data->alias = &pdev->dev;
+
+       atomic_set(&dev_data->bind, 0);
+
+       dev->archdata.iommu = dev_data;
+
+
+       return 0;
+}
+
+static void iommu_uninit_device(struct device *dev)
+{
+       kfree(dev->archdata.iommu);
+}
+
+void __init amd_iommu_uninit_devices(void)
+{
+       struct pci_dev *pdev = NULL;
+
+       for_each_pci_dev(pdev) {
+
+               if (!check_device(&pdev->dev))
+                       continue;
+
+               iommu_uninit_device(&pdev->dev);
+       }
+}
+
+int __init amd_iommu_init_devices(void)
+{
+       struct pci_dev *pdev = NULL;
+       int ret = 0;
+
+       for_each_pci_dev(pdev) {
+
+               if (!check_device(&pdev->dev))
+                       continue;
+
+               ret = iommu_init_device(&pdev->dev);
+               if (ret)
+                       goto out_free;
+       }
+
+       return 0;
+
+out_free:
+
+       amd_iommu_uninit_devices();
+
+       return ret;
+}
 #ifdef CONFIG_AMD_IOMMU_STATS
 
 /*
@@ -158,7 +223,6 @@ DECLARE_STATS_COUNTER(alloced_io_mem);
 DECLARE_STATS_COUNTER(total_map_requests);
 
 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)
@@ -176,9 +240,6 @@ static void amd_iommu_stats_init(void)
        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);
 
@@ -261,6 +322,7 @@ static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
                break;
        case EVENT_TYPE_ILL_CMD:
                printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
+               iommu->reset_in_progress = true;
                reset_iommu_command_buffer(iommu);
                dump_command(address);
                break;
@@ -383,11 +445,8 @@ static void __iommu_wait_for_completion(struct amd_iommu *iommu)
        status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
        writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
 
-       if (unlikely(i == EXIT_LOOP_COUNT)) {
-               spin_unlock(&iommu->lock);
-               reset_iommu_command_buffer(iommu);
-               spin_lock(&iommu->lock);
-       }
+       if (unlikely(i == EXIT_LOOP_COUNT))
+               iommu->reset_in_progress = true;
 }
 
 /*
@@ -434,6 +493,9 @@ static int iommu_completion_wait(struct amd_iommu *iommu)
 out:
        spin_unlock_irqrestore(&iommu->lock, flags);
 
+       if (iommu->reset_in_progress)
+               reset_iommu_command_buffer(iommu);
+
        return 0;
 }
 
@@ -456,20 +518,21 @@ static void iommu_flush_complete(struct protection_domain *domain)
 /*
  * Command send function for invalidating a device table entry
  */
-static int iommu_queue_inv_dev_entry(struct amd_iommu *iommu, u16 devid)
+static int iommu_flush_device(struct device *dev)
 {
+       struct amd_iommu *iommu;
        struct iommu_cmd cmd;
-       int ret;
+       u16 devid;
 
-       BUG_ON(iommu == NULL);
+       devid = get_device_id(dev);
+       iommu = amd_iommu_rlookup_table[devid];
 
+       /* Build command */
        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;
+       return iommu_queue_command(iommu, &cmd);
 }
 
 static void __iommu_build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
@@ -559,30 +622,43 @@ static void iommu_flush_tlb_pde(struct protection_domain *domain)
        __iommu_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
 }
 
+
 /*
- * This function flushes all domains that have devices on the given IOMMU
+ * This function flushes the DTEs for all devices in domain
  */
-static void flush_all_domains_on_iommu(struct amd_iommu *iommu)
+static void iommu_flush_domain_devices(struct protection_domain *domain)
+{
+       struct iommu_dev_data *dev_data;
+       unsigned long flags;
+
+       spin_lock_irqsave(&domain->lock, flags);
+
+       list_for_each_entry(dev_data, &domain->dev_list, list)
+               iommu_flush_device(dev_data->dev);
+
+       spin_unlock_irqrestore(&domain->lock, flags);
+}
+
+static void iommu_flush_all_domain_devices(void)
 {
-       u64 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
        struct protection_domain *domain;
        unsigned long flags;
 
        spin_lock_irqsave(&amd_iommu_pd_lock, flags);
 
        list_for_each_entry(domain, &amd_iommu_pd_list, list) {
-               if (domain->dev_iommu[iommu->index] == 0)
-                       continue;
-
-               spin_lock(&domain->lock);
-               iommu_queue_inv_iommu_pages(iommu, address, domain->id, 1, 1);
+               iommu_flush_domain_devices(domain);
                iommu_flush_complete(domain);
-               spin_unlock(&domain->lock);
        }
 
        spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
 }
 
+void amd_iommu_flush_all_devices(void)
+{
+       iommu_flush_all_domain_devices();
+}
+
 /*
  * This function uses heavy locking and may disable irqs for some time. But
  * this is no issue because it is only called during resume.
@@ -604,65 +680,120 @@ void amd_iommu_flush_all_domains(void)
        spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
 }
 
-static void flush_all_devices_for_iommu(struct amd_iommu *iommu)
+static void reset_iommu_command_buffer(struct amd_iommu *iommu)
 {
-       int i;
+       pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
 
-       for (i = 0; i <= amd_iommu_last_bdf; ++i) {
-               if (iommu != amd_iommu_rlookup_table[i])
-                       continue;
+       if (iommu->reset_in_progress)
+               panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
 
-               iommu_queue_inv_dev_entry(iommu, i);
-               iommu_completion_wait(iommu);
-       }
+       amd_iommu_reset_cmd_buffer(iommu);
+       amd_iommu_flush_all_devices();
+       amd_iommu_flush_all_domains();
+
+       iommu->reset_in_progress = false;
 }
 
-static void flush_devices_by_domain(struct protection_domain *domain)
+/****************************************************************************
+ *
+ * The functions below are used the create the page table mappings for
+ * unity mapped regions.
+ *
+ ****************************************************************************/
+
+/*
+ * This function is used to add another level to an IO page table. Adding
+ * another level increases the size of the address space by 9 bits to a size up
+ * to 64 bits.
+ */
+static bool increase_address_space(struct protection_domain *domain,
+                                  gfp_t gfp)
 {
-       struct amd_iommu *iommu;
-       int i;
+       u64 *pte;
 
-       for (i = 0; i <= amd_iommu_last_bdf; ++i) {
-               if ((domain == NULL && amd_iommu_pd_table[i] == NULL) ||
-                   (amd_iommu_pd_table[i] != domain))
-                       continue;
+       if (domain->mode == PAGE_MODE_6_LEVEL)
+               /* address space already 64 bit large */
+               return false;
 
-               iommu = amd_iommu_rlookup_table[i];
-               if (!iommu)
-                       continue;
+       pte = (void *)get_zeroed_page(gfp);
+       if (!pte)
+               return false;
 
-               iommu_queue_inv_dev_entry(iommu, i);
-               iommu_completion_wait(iommu);
-       }
+       *pte             = PM_LEVEL_PDE(domain->mode,
+                                       virt_to_phys(domain->pt_root));
+       domain->pt_root  = pte;
+       domain->mode    += 1;
+       domain->updated  = true;
+
+       return true;
 }
 
-static void reset_iommu_command_buffer(struct amd_iommu *iommu)
+static u64 *alloc_pte(struct protection_domain *domain,
+                     unsigned long address,
+                     int end_lvl,
+                     u64 **pte_page,
+                     gfp_t gfp)
 {
-       pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
+       u64 *pte, *page;
+       int level;
 
-       if (iommu->reset_in_progress)
-               panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
+       while (address > PM_LEVEL_SIZE(domain->mode))
+               increase_address_space(domain, gfp);
 
-       iommu->reset_in_progress = true;
+       level =  domain->mode - 1;
+       pte   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
 
-       amd_iommu_reset_cmd_buffer(iommu);
-       flush_all_devices_for_iommu(iommu);
-       flush_all_domains_on_iommu(iommu);
+       while (level > end_lvl) {
+               if (!IOMMU_PTE_PRESENT(*pte)) {
+                       page = (u64 *)get_zeroed_page(gfp);
+                       if (!page)
+                               return NULL;
+                       *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
+               }
 
-       iommu->reset_in_progress = false;
+               level -= 1;
+
+               pte = IOMMU_PTE_PAGE(*pte);
+
+               if (pte_page && level == end_lvl)
+                       *pte_page = pte;
+
+               pte = &pte[PM_LEVEL_INDEX(level, address)];
+       }
+
+       return pte;
 }
 
-void amd_iommu_flush_all_devices(void)
+/*
+ * This function checks if there is a PTE for a given dma address. If
+ * there is one, it returns the pointer to it.
+ */
+static u64 *fetch_pte(struct protection_domain *domain,
+                     unsigned long address, int map_size)
 {
-       flush_devices_by_domain(NULL);
-}
+       int level;
+       u64 *pte;
 
-/****************************************************************************
- *
- * The functions below are used the create the page table mappings for
- * unity mapped regions.
- *
- ****************************************************************************/
+       level =  domain->mode - 1;
+       pte   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
+
+       while (level > map_size) {
+               if (!IOMMU_PTE_PRESENT(*pte))
+                       return NULL;
+
+               level -= 1;
+
+               pte = IOMMU_PTE_PAGE(*pte);
+               pte = &pte[PM_LEVEL_INDEX(level, address)];
+
+               if ((PM_PTE_LEVEL(*pte) == 0) && level != map_size) {
+                       pte = NULL;
+                       break;
+               }
+       }
+
+       return pte;
+}
 
 /*
  * Generic mapping functions. It maps a physical address into a DMA
@@ -734,28 +865,6 @@ static int iommu_for_unity_map(struct amd_iommu *iommu,
 }
 
 /*
- * 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.
  */
@@ -784,6 +893,28 @@ static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
 }
 
 /*
+ * 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;
+}
+
+/*
  * Inits the unity mappings required for a specific device
  */
 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
@@ -820,34 +951,23 @@ static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
  */
 
 /*
- * This function checks if there is a PTE for a given dma address. If
- * there is one, it returns the pointer to it.
+ * Used to reserve address ranges in the aperture (e.g. for exclusion
+ * ranges.
  */
-static u64 *fetch_pte(struct protection_domain *domain,
-                     unsigned long address, int map_size)
+static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
+                                     unsigned long start_page,
+                                     unsigned int pages)
 {
-       int level;
-       u64 *pte;
-
-       level =  domain->mode - 1;
-       pte   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
-
-       while (level > map_size) {
-               if (!IOMMU_PTE_PRESENT(*pte))
-                       return NULL;
-
-               level -= 1;
+       unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
 
-               pte = IOMMU_PTE_PAGE(*pte);
-               pte = &pte[PM_LEVEL_INDEX(level, address)];
+       if (start_page + pages > last_page)
+               pages = last_page - start_page;
 
-               if ((PM_PTE_LEVEL(*pte) == 0) && level != map_size) {
-                       pte = NULL;
-                       break;
-               }
+       for (i = start_page; i < start_page + pages; ++i) {
+               int index = i / APERTURE_RANGE_PAGES;
+               int page  = i % APERTURE_RANGE_PAGES;
+               __set_bit(page, dom->aperture[index]->bitmap);
        }
-
-       return pte;
 }
 
 /*
@@ -860,7 +980,7 @@ static int alloc_new_range(struct dma_ops_domain *dma_dom,
 {
        int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
        struct amd_iommu *iommu;
-       int i;
+       unsigned long i;
 
 #ifdef CONFIG_IOMMU_STRESS
        populate = false;
@@ -1042,7 +1162,7 @@ static void dma_ops_free_addresses(struct dma_ops_domain *dom,
 
        address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
 
-       iommu_area_free(range->bitmap, address, pages);
+       bitmap_clear(range->bitmap, address, pages);
 
 }
 
@@ -1108,26 +1228,6 @@ static void domain_id_free(int id)
        write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
 }
 
-/*
- * 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 i, last_page = dom->aperture_size >> PAGE_SHIFT;
-
-       if (start_page + pages > last_page)
-               pages = last_page - start_page;
-
-       for (i = start_page; i < start_page + pages; ++i) {
-               int index = i / APERTURE_RANGE_PAGES;
-               int page  = i % APERTURE_RANGE_PAGES;
-               __set_bit(page, dom->aperture[index]->bitmap);
-       }
-}
-
 static void free_pagetable(struct protection_domain *domain)
 {
        int i, j;
@@ -1201,6 +1301,7 @@ static struct dma_ops_domain *dma_ops_domain_alloc(void)
        dma_dom->domain.id = domain_id_alloc();
        if (dma_dom->domain.id == 0)
                goto free_dma_dom;
+       INIT_LIST_HEAD(&dma_dom->domain.dev_list);
        dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
        dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
        dma_dom->domain.flags = PD_DMA_OPS_MASK;
@@ -1243,11 +1344,8 @@ static bool dma_ops_domain(struct protection_domain *domain)
 
 static void set_dte_entry(u16 devid, struct protection_domain *domain)
 {
-       struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
        u64 pte_root = virt_to_phys(domain->pt_root);
 
-       BUG_ON(amd_iommu_pd_table[devid] != NULL);
-
        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;
@@ -1255,39 +1353,62 @@ static void set_dte_entry(u16 devid, struct protection_domain *domain)
        amd_iommu_dev_table[devid].data[2] = domain->id;
        amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
        amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
-
-       amd_iommu_pd_table[devid] = domain;
-
-       /* Do reference counting */
-       domain->dev_iommu[iommu->index] += 1;
-       domain->dev_cnt                 += 1;
-
-       /* Flush the changes DTE entry */
-       iommu_queue_inv_dev_entry(iommu, devid);
 }
 
 static void clear_dte_entry(u16 devid)
 {
-       struct protection_domain *domain = amd_iommu_pd_table[devid];
-       struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
-
-       BUG_ON(domain == NULL);
-
-       /* 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;
 
        amd_iommu_apply_erratum_63(devid);
+}
+
+static void do_attach(struct device *dev, struct protection_domain *domain)
+{
+       struct iommu_dev_data *dev_data;
+       struct amd_iommu *iommu;
+       u16 devid;
+
+       devid    = get_device_id(dev);
+       iommu    = amd_iommu_rlookup_table[devid];
+       dev_data = get_dev_data(dev);
+
+       /* Update data structures */
+       dev_data->domain = domain;
+       list_add(&dev_data->list, &domain->dev_list);
+       set_dte_entry(devid, domain);
+
+       /* Do reference counting */
+       domain->dev_iommu[iommu->index] += 1;
+       domain->dev_cnt                 += 1;
+
+       /* Flush the DTE entry */
+       iommu_flush_device(dev);
+}
+
+static void do_detach(struct device *dev)
+{
+       struct iommu_dev_data *dev_data;
+       struct amd_iommu *iommu;
+       u16 devid;
+
+       devid    = get_device_id(dev);
+       iommu    = amd_iommu_rlookup_table[devid];
+       dev_data = get_dev_data(dev);
 
        /* decrease reference counters */
-       domain->dev_iommu[iommu->index] -= 1;
-       domain->dev_cnt                 -= 1;
+       dev_data->domain->dev_iommu[iommu->index] -= 1;
+       dev_data->domain->dev_cnt                 -= 1;
 
-       iommu_queue_inv_dev_entry(iommu, devid);
+       /* Update data structures */
+       dev_data->domain = NULL;
+       list_del(&dev_data->list);
+       clear_dte_entry(devid);
+
+       /* Flush the DTE entry */
+       iommu_flush_device(dev);
 }
 
 /*
@@ -1297,28 +1418,39 @@ static void clear_dte_entry(u16 devid)
 static int __attach_device(struct device *dev,
                           struct protection_domain *domain)
 {
-       u16 devid = get_device_id(dev);
-       u16 alias = amd_iommu_alias_table[devid];
+       struct iommu_dev_data *dev_data, *alias_data;
+
+       dev_data   = get_dev_data(dev);
+       alias_data = get_dev_data(dev_data->alias);
+
+       if (!alias_data)
+               return -EINVAL;
 
        /* lock domain */
        spin_lock(&domain->lock);
 
        /* Some sanity checks */
-       if (amd_iommu_pd_table[alias] != NULL &&
-           amd_iommu_pd_table[alias] != domain)
+       if (alias_data->domain != NULL &&
+           alias_data->domain != domain)
                return -EBUSY;
 
-       if (amd_iommu_pd_table[devid] != NULL &&
-           amd_iommu_pd_table[devid] != domain)
+       if (dev_data->domain != NULL &&
+           dev_data->domain != domain)
                return -EBUSY;
 
        /* Do real assignment */
-       if (alias != devid &&
-           amd_iommu_pd_table[alias] == NULL)
-               set_dte_entry(alias, domain);
+       if (dev_data->alias != dev) {
+               alias_data = get_dev_data(dev_data->alias);
+               if (alias_data->domain == NULL)
+                       do_attach(dev_data->alias, domain);
 
-       if (amd_iommu_pd_table[devid] == NULL)
-               set_dte_entry(devid, domain);
+               atomic_inc(&alias_data->bind);
+       }
+
+       if (dev_data->domain == NULL)
+               do_attach(dev, domain);
+
+       atomic_inc(&dev_data->bind);
 
        /* ready */
        spin_unlock(&domain->lock);
@@ -1355,18 +1487,35 @@ static int attach_device(struct device *dev,
  */
 static void __detach_device(struct device *dev)
 {
-       u16 devid = get_device_id(dev);
-       struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
+       struct iommu_dev_data *dev_data = get_dev_data(dev);
+       struct iommu_dev_data *alias_data;
+       struct protection_domain *domain;
+       unsigned long flags;
 
-       BUG_ON(!iommu);
+       BUG_ON(!dev_data->domain);
 
-       clear_dte_entry(devid);
+       domain = dev_data->domain;
+
+       spin_lock_irqsave(&domain->lock, flags);
+
+       if (dev_data->alias != dev) {
+               alias_data = get_dev_data(dev_data->alias);
+               if (atomic_dec_and_test(&alias_data->bind))
+                       do_detach(dev_data->alias);
+       }
+
+       if (atomic_dec_and_test(&dev_data->bind))
+               do_detach(dev);
+
+       spin_unlock_irqrestore(&domain->lock, flags);
 
        /*
         * If we run in passthrough mode the device must be assigned to the
-        * passthrough domain if it is detached from any other domain
+        * passthrough domain if it is detached from any other domain.
+        * Make sure we can deassign from the pt_domain itself.
         */
-       if (iommu_pass_through)
+       if (iommu_pass_through &&
+           (dev_data->domain == NULL && domain != pt_domain))
                __attach_device(dev, pt_domain);
 }
 
@@ -1390,18 +1539,23 @@ static void detach_device(struct device *dev)
 static struct protection_domain *domain_for_device(struct device *dev)
 {
        struct protection_domain *dom;
+       struct iommu_dev_data *dev_data, *alias_data;
        unsigned long flags;
        u16 devid, alias;
 
-       devid = get_device_id(dev);
-       alias = amd_iommu_alias_table[devid];
+       devid      = get_device_id(dev);
+       alias      = amd_iommu_alias_table[devid];
+       dev_data   = get_dev_data(dev);
+       alias_data = get_dev_data(dev_data->alias);
+       if (!alias_data)
+               return NULL;
 
        read_lock_irqsave(&amd_iommu_devtable_lock, flags);
-       dom = amd_iommu_pd_table[devid];
+       dom = dev_data->domain;
        if (dom == NULL &&
-           amd_iommu_pd_table[alias] != NULL) {
-               __attach_device(dev, amd_iommu_pd_table[alias]);
-               dom = amd_iommu_pd_table[devid];
+           alias_data->domain != NULL) {
+               __attach_device(dev, alias_data->domain);
+               dom = alias_data->domain;
        }
 
        read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
@@ -1424,14 +1578,12 @@ static int device_change_notifier(struct notifier_block *nb,
 
        devid  = get_device_id(dev);
        iommu  = amd_iommu_rlookup_table[devid];
-       domain = domain_for_device(dev);
-
-       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_UNBOUND_DRIVER:
+
+               domain = domain_for_device(dev);
+
                if (!domain)
                        goto out;
                if (iommu_pass_through)
@@ -1439,6 +1591,11 @@ static int device_change_notifier(struct notifier_block *nb,
                detach_device(dev);
                break;
        case BUS_NOTIFY_ADD_DEVICE:
+
+               iommu_init_device(dev);
+
+               domain = domain_for_device(dev);
+
                /* allocate a protection domain if a device is added */
                dma_domain = find_protection_domain(devid);
                if (dma_domain)
@@ -1453,11 +1610,15 @@ static int device_change_notifier(struct notifier_block *nb,
                spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
 
                break;
+       case BUS_NOTIFY_DEL_DEVICE:
+
+               iommu_uninit_device(dev);
+
        default:
                goto out;
        }
 
-       iommu_queue_inv_dev_entry(iommu, devid);
+       iommu_flush_device(dev);
        iommu_completion_wait(iommu);
 
 out:
@@ -1468,6 +1629,11 @@ static struct notifier_block device_nb = {
        .notifier_call = device_change_notifier,
 };
 
+void amd_iommu_init_notifier(void)
+{
+       bus_register_notifier(&pci_bus_type, &device_nb);
+}
+
 /*****************************************************************************
  *
  * The next functions belong to the dma_ops mapping/unmapping code.
@@ -1510,15 +1676,11 @@ static struct protection_domain *get_domain(struct device *dev)
 
 static void update_device_table(struct protection_domain *domain)
 {
-       unsigned long flags;
-       int i;
+       struct iommu_dev_data *dev_data;
 
-       for (i = 0; i <= amd_iommu_last_bdf; ++i) {
-               if (amd_iommu_pd_table[i] != domain)
-                       continue;
-               write_lock_irqsave(&amd_iommu_devtable_lock, flags);
-               set_dte_entry(i, domain);
-               write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
+       list_for_each_entry(dev_data, &domain->dev_list, list) {
+               u16 devid = get_device_id(dev_data->dev);
+               set_dte_entry(devid, domain);
        }
 }
 
@@ -1528,76 +1690,13 @@ static void update_domain(struct protection_domain *domain)
                return;
 
        update_device_table(domain);
-       flush_devices_by_domain(domain);
+       iommu_flush_domain_devices(domain);
        iommu_flush_tlb_pde(domain);
 
        domain->updated = false;
 }
 
 /*
- * This function is used to add another level to an IO page table. Adding
- * another level increases the size of the address space by 9 bits to a size up
- * to 64 bits.
- */
-static bool increase_address_space(struct protection_domain *domain,
-                                  gfp_t gfp)
-{
-       u64 *pte;
-
-       if (domain->mode == PAGE_MODE_6_LEVEL)
-               /* address space already 64 bit large */
-               return false;
-
-       pte = (void *)get_zeroed_page(gfp);
-       if (!pte)
-               return false;
-
-       *pte             = PM_LEVEL_PDE(domain->mode,
-                                       virt_to_phys(domain->pt_root));
-       domain->pt_root  = pte;
-       domain->mode    += 1;
-       domain->updated  = true;
-
-       return true;
-}
-
-static u64 *alloc_pte(struct protection_domain *domain,
-                     unsigned long address,
-                     int end_lvl,
-                     u64 **pte_page,
-                     gfp_t gfp)
-{
-       u64 *pte, *page;
-       int level;
-
-       while (address > PM_LEVEL_SIZE(domain->mode))
-               increase_address_space(domain, gfp);
-
-       level =  domain->mode - 1;
-       pte   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
-
-       while (level > end_lvl) {
-               if (!IOMMU_PTE_PRESENT(*pte)) {
-                       page = (u64 *)get_zeroed_page(gfp);
-                       if (!page)
-                               return NULL;
-                       *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
-               }
-
-               level -= 1;
-
-               pte = IOMMU_PTE_PAGE(*pte);
-
-               if (pte_page && level == end_lvl)
-                       *pte_page = pte;
-
-               pte = &pte[PM_LEVEL_INDEX(level, address)];
-       }
-
-       return pte;
-}
-
-/*
  * This function fetches the PTE for a given address in the aperture
  */
 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
@@ -1731,7 +1830,7 @@ retry:
                        goto out;
 
                /*
-                * aperture was sucessfully enlarged by 128 MB, try
+                * aperture was successfully enlarged by 128 MB, try
                 * allocation again
                 */
                goto retry;
@@ -2124,6 +2223,12 @@ static struct dma_map_ops amd_iommu_dma_ops = {
 /*
  * The function which clues the AMD IOMMU driver into dma_ops.
  */
+
+void __init amd_iommu_init_api(void)
+{
+       register_iommu(&amd_iommu_ops);
+}
+
 int __init amd_iommu_init_dma_ops(void)
 {
        struct amd_iommu *iommu;
@@ -2145,11 +2250,9 @@ int __init amd_iommu_init_dma_ops(void)
        }
 
        /*
-        * If device isolation is enabled, pre-allocate the protection
-        * domains for each device.
+        * Pre-allocate the protection domains for each device.
         */
-       if (amd_iommu_isolate)
-               prealloc_protection_domains();
+       prealloc_protection_domains();
 
        iommu_detected = 1;
        swiotlb = 0;
@@ -2161,10 +2264,6 @@ int __init amd_iommu_init_dma_ops(void)
        /* Make the driver finally visible to the drivers */
        dma_ops = &amd_iommu_dma_ops;
 
-       register_iommu(&amd_iommu_ops);
-
-       bus_register_notifier(&pci_bus_type, &device_nb);
-
        amd_iommu_stats_init();
 
        return 0;
@@ -2191,14 +2290,17 @@ free_domains:
 
 static void cleanup_domain(struct protection_domain *domain)
 {
+       struct iommu_dev_data *dev_data, *next;
        unsigned long flags;
-       u16 devid;
 
        write_lock_irqsave(&amd_iommu_devtable_lock, flags);
 
-       for (devid = 0; devid <= amd_iommu_last_bdf; ++devid)
-               if (amd_iommu_pd_table[devid] == domain)
-                       clear_dte_entry(devid);
+       list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
+               struct device *dev = dev_data->dev;
+
+               do_detach(dev);
+               atomic_set(&dev_data->bind, 0);
+       }
 
        write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
 }
@@ -2228,6 +2330,7 @@ static struct protection_domain *protection_domain_alloc(void)
        domain->id = domain_id_alloc();
        if (!domain->id)
                goto out_err;
+       INIT_LIST_HEAD(&domain->dev_list);
 
        add_domain_to_list(domain);
 
@@ -2286,6 +2389,7 @@ static void amd_iommu_domain_destroy(struct iommu_domain *dom)
 static void amd_iommu_detach_device(struct iommu_domain *dom,
                                    struct device *dev)
 {
+       struct iommu_dev_data *dev_data = dev->archdata.iommu;
        struct amd_iommu *iommu;
        u16 devid;
 
@@ -2294,14 +2398,14 @@ static void amd_iommu_detach_device(struct iommu_domain *dom,
 
        devid = get_device_id(dev);
 
-       if (amd_iommu_pd_table[devid] != NULL)
+       if (dev_data->domain != NULL)
                detach_device(dev);
 
        iommu = amd_iommu_rlookup_table[devid];
        if (!iommu)
                return;
 
-       iommu_queue_inv_dev_entry(iommu, devid);
+       iommu_flush_device(dev);
        iommu_completion_wait(iommu);
 }
 
@@ -2309,7 +2413,7 @@ static int amd_iommu_attach_device(struct iommu_domain *dom,
                                   struct device *dev)
 {
        struct protection_domain *domain = dom->priv;
-       struct protection_domain *old_domain;
+       struct iommu_dev_data *dev_data;
        struct amd_iommu *iommu;
        int ret;
        u16 devid;
@@ -2317,14 +2421,15 @@ static int amd_iommu_attach_device(struct iommu_domain *dom,
        if (!check_device(dev))
                return -EINVAL;
 
+       dev_data = dev->archdata.iommu;
+
        devid = get_device_id(dev);
 
        iommu = amd_iommu_rlookup_table[devid];
        if (!iommu)
                return -EINVAL;
 
-       old_domain = amd_iommu_pd_table[devid];
-       if (old_domain)
+       if (dev_data->domain)
                detach_device(dev);
 
        ret = attach_device(dev, domain);
@@ -2432,7 +2537,7 @@ int __init amd_iommu_init_passthrough(void)
        struct pci_dev *dev = NULL;
        u16 devid;
 
-       /* allocate passthroug domain */
+       /* allocate passthrough domain */
        pt_domain = protection_domain_alloc();
        if (!pt_domain)
                return -ENOMEM;