AMD IOMMU: add stats counter for map_single requests
[pandora-kernel.git] / arch / x86 / kernel / amd_iommu.c
index 15456a3..ef37786 100644 (file)
 #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>
@@ -38,6 +42,10 @@ static DEFINE_RWLOCK(amd_iommu_devtable_lock);
 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
  */
@@ -47,6 +55,48 @@ struct iommu_cmd {
 
 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);
+DECLARE_STATS_COUNTER(cnt_map_single);
+
+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);
+       amd_iommu_stats_add(&cnt_map_single);
+}
+
+#endif
 
 /* returns !0 if the IOMMU is caching non-present entries in its TLB */
 static int iommu_has_npcache(struct amd_iommu *iommu)
@@ -189,7 +239,7 @@ static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
        spin_lock_irqsave(&iommu->lock, flags);
        ret = __iommu_queue_command(iommu, cmd);
        if (!ret)
-               iommu->need_sync = 1;
+               iommu->need_sync = true;
        spin_unlock_irqrestore(&iommu->lock, flags);
 
        return ret;
@@ -205,6 +255,8 @@ static void __iommu_wait_for_completion(struct amd_iommu *iommu)
        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 */
@@ -254,7 +306,7 @@ static int iommu_completion_wait(struct amd_iommu *iommu)
 
        ret = __iommu_completion_wait(iommu);
 
-       iommu->need_sync = 0;
+       iommu->need_sync = false;
 
        if (ret)
                goto out;
@@ -440,6 +492,30 @@ static int iommu_map_page(struct protection_domain *dom,
        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.
@@ -844,7 +920,6 @@ static void attach_device(struct amd_iommu *iommu,
        iommu_queue_inv_dev_entry(iommu, devid);
 }
 
-#ifdef CONFIG_IOMMU_API
 /*
  * Removes a device from a protection domain (unlocked)
  */
@@ -881,7 +956,79 @@ static void detach_device(struct protection_domain *domain, u16 devid)
        __detach_device(domain, devid);
        write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
 }
-#endif
+
+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;
+       case BUS_NOTIFY_ADD_DEVICE:
+               /* allocate a protection domain if a device is added */
+               dma_domain = find_protection_domain(devid);
+               if (dma_domain)
+                       goto out;
+               dma_domain = dma_ops_domain_alloc(iommu, order);
+               if (!dma_domain)
+                       goto out;
+               dma_domain->target_dev = devid;
+
+               spin_lock_irqsave(&iommu_pd_list_lock, flags);
+               list_add_tail(&dma_domain->list, &iommu_pd_list);
+               spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
+
+               break;
+       default:
+               goto out;
+       }
+
+       iommu_queue_inv_dev_entry(iommu, devid);
+       iommu_completion_wait(iommu);
+
+out:
+       return 0;
+}
+
+struct notifier_block device_nb = {
+       .notifier_call = device_change_notifier,
+};
 
 /*****************************************************************************
  *
@@ -970,8 +1117,7 @@ static int get_device_resources(struct device *dev,
                *domain = &dma_dom->domain;
                attach_device(*iommu, *domain, *bdf);
                printk(KERN_INFO "AMD IOMMU: Using protection domain %d for "
-                               "device ", (*domain)->id);
-               print_devid(_bdf, 1);
+                               "device %s\n", (*domain)->id, dev_name(dev));
        }
 
        if (domain_for_device(_bdf) == NULL)
@@ -1134,6 +1280,8 @@ static dma_addr_t map_single(struct device *dev, phys_addr_t paddr,
        dma_addr_t addr;
        u64 dma_mask;
 
+       INC_STATS_COUNTER(cnt_map_single);
+
        if (!check_device(dev))
                return bad_dma_address;
 
@@ -1440,7 +1588,7 @@ void prealloc_protection_domains(void)
        u16 devid;
 
        while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
-               devid = (dev->bus->number << 8) | dev->devfn;
+               devid = calc_devid(dev->bus->number, dev->devfn);
                if (devid > amd_iommu_last_bdf)
                        continue;
                devid = amd_iommu_alias_table[devid];
@@ -1487,6 +1635,7 @@ int __init amd_iommu_init_dma_ops(void)
                iommu->default_dom = dma_ops_domain_alloc(iommu, order);
                if (iommu->default_dom == NULL)
                        return -ENOMEM;
+               iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
                ret = iommu_init_unity_mappings(iommu);
                if (ret)
                        goto free_domains;
@@ -1510,6 +1659,14 @@ int __init amd_iommu_init_dma_ops(void)
        /* Make the driver finally visible to the drivers */
        dma_ops = &amd_iommu_dma_ops;
 
+#ifdef CONFIG_IOMMU_API
+       register_iommu(&amd_iommu_ops);
+#endif
+
+       bus_register_notifier(&pci_bus_type, &device_nb);
+
+       amd_iommu_stats_init();
+
        return 0;
 
 free_domains:
@@ -1521,3 +1678,227 @@ free_domains:
 
        return ret;
 }
+
+/*****************************************************************************
+ *
+ * The following functions belong to the exported interface of AMD IOMMU
+ *
+ * This interface allows access to lower level functions of the IOMMU
+ * like protection domain handling and assignement of devices to domains
+ * which is not possible with the dma_ops interface.
+ *
+ *****************************************************************************/
+
+#ifdef CONFIG_IOMMU_API
+
+static void cleanup_domain(struct protection_domain *domain)
+{
+       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)
+                       __detach_device(domain, devid);
+
+       write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
+}
+
+static int amd_iommu_domain_init(struct iommu_domain *dom)
+{
+       struct protection_domain *domain;
+
+       domain = kzalloc(sizeof(*domain), GFP_KERNEL);
+       if (!domain)
+               return -ENOMEM;
+
+       spin_lock_init(&domain->lock);
+       domain->mode = PAGE_MODE_3_LEVEL;
+       domain->id = domain_id_alloc();
+       if (!domain->id)
+               goto out_free;
+       domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
+       if (!domain->pt_root)
+               goto out_free;
+
+       dom->priv = domain;
+
+       return 0;
+
+out_free:
+       kfree(domain);
+
+       return -ENOMEM;
+}
+
+static void amd_iommu_domain_destroy(struct iommu_domain *dom)
+{
+       struct protection_domain *domain = dom->priv;
+
+       if (!domain)
+               return;
+
+       if (domain->dev_cnt > 0)
+               cleanup_domain(domain);
+
+       BUG_ON(domain->dev_cnt != 0);
+
+       free_pagetable(domain);
+
+       domain_id_free(domain->id);
+
+       kfree(domain);
+
+       dom->priv = NULL;
+}
+
+static void amd_iommu_detach_device(struct iommu_domain *dom,
+                                   struct device *dev)
+{
+       struct protection_domain *domain = dom->priv;
+       struct amd_iommu *iommu;
+       struct pci_dev *pdev;
+       u16 devid;
+
+       if (dev->bus != &pci_bus_type)
+               return;
+
+       pdev = to_pci_dev(dev);
+
+       devid = calc_devid(pdev->bus->number, pdev->devfn);
+
+       if (devid > 0)
+               detach_device(domain, devid);
+
+       iommu = amd_iommu_rlookup_table[devid];
+       if (!iommu)
+               return;
+
+       iommu_queue_inv_dev_entry(iommu, devid);
+       iommu_completion_wait(iommu);
+}
+
+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 amd_iommu *iommu;
+       struct pci_dev *pdev;
+       u16 devid;
+
+       if (dev->bus != &pci_bus_type)
+               return -EINVAL;
+
+       pdev = to_pci_dev(dev);
+
+       devid = calc_devid(pdev->bus->number, pdev->devfn);
+
+       if (devid >= amd_iommu_last_bdf ||
+                       devid != amd_iommu_alias_table[devid])
+               return -EINVAL;
+
+       iommu = amd_iommu_rlookup_table[devid];
+       if (!iommu)
+               return -EINVAL;
+
+       old_domain = domain_for_device(devid);
+       if (old_domain)
+               return -EBUSY;
+
+       attach_device(iommu, domain, devid);
+
+       iommu_completion_wait(iommu);
+
+       return 0;
+}
+
+static int amd_iommu_map_range(struct iommu_domain *dom,
+                              unsigned long iova, phys_addr_t paddr,
+                              size_t size, int iommu_prot)
+{
+       struct protection_domain *domain = dom->priv;
+       unsigned long i,  npages = iommu_num_pages(paddr, size, PAGE_SIZE);
+       int prot = 0;
+       int ret;
+
+       if (iommu_prot & IOMMU_READ)
+               prot |= IOMMU_PROT_IR;
+       if (iommu_prot & IOMMU_WRITE)
+               prot |= IOMMU_PROT_IW;
+
+       iova  &= PAGE_MASK;
+       paddr &= PAGE_MASK;
+
+       for (i = 0; i < npages; ++i) {
+               ret = iommu_map_page(domain, iova, paddr, prot);
+               if (ret)
+                       return ret;
+
+               iova  += PAGE_SIZE;
+               paddr += PAGE_SIZE;
+       }
+
+       return 0;
+}
+
+static void amd_iommu_unmap_range(struct iommu_domain *dom,
+                                 unsigned long iova, size_t size)
+{
+
+       struct protection_domain *domain = dom->priv;
+       unsigned long i,  npages = iommu_num_pages(iova, size, PAGE_SIZE);
+
+       iova  &= PAGE_MASK;
+
+       for (i = 0; i < npages; ++i) {
+               iommu_unmap_page(domain, iova);
+               iova  += PAGE_SIZE;
+       }
+
+       iommu_flush_domain(domain->id);
+}
+
+static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
+                                         unsigned long iova)
+{
+       struct protection_domain *domain = dom->priv;
+       unsigned long offset = iova & ~PAGE_MASK;
+       phys_addr_t paddr;
+       u64 *pte;
+
+       pte = &domain->pt_root[IOMMU_PTE_L2_INDEX(iova)];
+
+       if (!IOMMU_PTE_PRESENT(*pte))
+               return 0;
+
+       pte = IOMMU_PTE_PAGE(*pte);
+       pte = &pte[IOMMU_PTE_L1_INDEX(iova)];
+
+       if (!IOMMU_PTE_PRESENT(*pte))
+               return 0;
+
+       pte = IOMMU_PTE_PAGE(*pte);
+       pte = &pte[IOMMU_PTE_L0_INDEX(iova)];
+
+       if (!IOMMU_PTE_PRESENT(*pte))
+               return 0;
+
+       paddr  = *pte & IOMMU_PAGE_MASK;
+       paddr |= offset;
+
+       return paddr;
+}
+
+static struct iommu_ops amd_iommu_ops = {
+       .domain_init = amd_iommu_domain_init,
+       .domain_destroy = amd_iommu_domain_destroy,
+       .attach_dev = amd_iommu_attach_device,
+       .detach_dev = amd_iommu_detach_device,
+       .map = amd_iommu_map_range,
+       .unmap = amd_iommu_unmap_range,
+       .iova_to_phys = amd_iommu_iova_to_phys,
+};
+
+#endif