vfio: Type1 IOMMU implementation
authorAlex Williamson <alex.williamson@redhat.com>
Tue, 31 Jul 2012 14:16:23 +0000 (08:16 -0600)
committerAlex Williamson <alex.williamson@redhat.com>
Tue, 31 Jul 2012 14:16:23 +0000 (08:16 -0600)
This VFIO IOMMU backend is designed primarily for AMD-Vi and Intel
VT-d hardware, but is potentially usable by anything supporting
similar mapping functionality.  We arbitrarily call this a Type1
backend for lack of a better name.  This backend has no IOVA
or host memory mapping restrictions for the user and is optimized
for relatively static mappings.  Mapped areas are pinned into system
memory.

Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
drivers/vfio/Kconfig
drivers/vfio/Makefile
drivers/vfio/vfio.c
drivers/vfio/vfio_iommu_type1.c [new file with mode: 0644]
include/linux/vfio.h

index 9acb1e729bd6003848792c06c802d4f87b89d9b4..128b97910b8e448725df25958ee1aa60171c929e 100644 (file)
@@ -1,6 +1,12 @@
+config VFIO_IOMMU_TYPE1
+       tristate
+       depends on VFIO
+       default n
+
 menuconfig VFIO
        tristate "VFIO Non-Privileged userspace driver framework"
        depends on IOMMU_API
+       select VFIO_IOMMU_TYPE1 if X86
        help
          VFIO provides a framework for secure userspace device drivers.
          See Documentation/vfio.txt for more details.
index 7500a67a42a0e24a61ae19bb185423c2789bc983..2398d4a0e38b9d93f2bb7dae81d713b50b273751 100644 (file)
@@ -1 +1,3 @@
 obj-$(CONFIG_VFIO) += vfio.o
+obj-$(CONFIG_VFIO_IOMMU_TYPE1) += vfio_iommu_type1.o
+obj-$(CONFIG_VFIO_PCI) += pci/
index 052e310aed722a4130add7c9eeb8e3139b01a55c..9591e2b509d794e93e596fe8f2c1da8c11691205 100644 (file)
@@ -1376,6 +1376,13 @@ static int __init vfio_init(void)
 
        pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
 
+       /*
+        * Attempt to load known iommu-drivers.  This gives us a working
+        * environment without the user needing to explicitly load iommu
+        * drivers.
+        */
+       request_module_nowait("vfio_iommu_type1");
+
        return 0;
 
 err_groups_cdev:
diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
new file mode 100644 (file)
index 0000000..6f3fbc4
--- /dev/null
@@ -0,0 +1,753 @@
+/*
+ * VFIO: IOMMU DMA mapping support for Type1 IOMMU
+ *
+ * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
+ *     Author: Alex Williamson <alex.williamson@redhat.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.
+ *
+ * Derived from original vfio:
+ * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
+ * Author: Tom Lyon, pugs@cisco.com
+ *
+ * We arbitrarily define a Type1 IOMMU as one matching the below code.
+ * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
+ * VT-d, but that makes it harder to re-use as theoretically anyone
+ * implementing a similar IOMMU could make use of this.  We expect the
+ * IOMMU to support the IOMMU API and have few to no restrictions around
+ * the IOVA range that can be mapped.  The Type1 IOMMU is currently
+ * optimized for relatively static mappings of a userspace process with
+ * userpsace pages pinned into memory.  We also assume devices and IOMMU
+ * domains are PCI based as the IOMMU API is still centered around a
+ * device/bus interface rather than a group interface.
+ */
+
+#include <linux/compat.h>
+#include <linux/device.h>
+#include <linux/fs.h>
+#include <linux/iommu.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+#include <linux/pci.h>         /* pci_bus_type */
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+#include <linux/vfio.h>
+#include <linux/workqueue.h>
+
+#define DRIVER_VERSION  "0.2"
+#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
+#define DRIVER_DESC     "Type1 IOMMU driver for VFIO"
+
+static bool allow_unsafe_interrupts;
+module_param_named(allow_unsafe_interrupts,
+                  allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
+MODULE_PARM_DESC(allow_unsafe_interrupts,
+                "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
+
+struct vfio_iommu {
+       struct iommu_domain     *domain;
+       struct mutex            lock;
+       struct list_head        dma_list;
+       struct list_head        group_list;
+       bool                    cache;
+};
+
+struct vfio_dma {
+       struct list_head        next;
+       dma_addr_t              iova;           /* Device address */
+       unsigned long           vaddr;          /* Process virtual addr */
+       long                    npage;          /* Number of pages */
+       int                     prot;           /* IOMMU_READ/WRITE */
+};
+
+struct vfio_group {
+       struct iommu_group      *iommu_group;
+       struct list_head        next;
+};
+
+/*
+ * This code handles mapping and unmapping of user data buffers
+ * into DMA'ble space using the IOMMU
+ */
+
+#define NPAGE_TO_SIZE(npage)   ((size_t)(npage) << PAGE_SHIFT)
+
+struct vwork {
+       struct mm_struct        *mm;
+       long                    npage;
+       struct work_struct      work;
+};
+
+/* delayed decrement/increment for locked_vm */
+static void vfio_lock_acct_bg(struct work_struct *work)
+{
+       struct vwork *vwork = container_of(work, struct vwork, work);
+       struct mm_struct *mm;
+
+       mm = vwork->mm;
+       down_write(&mm->mmap_sem);
+       mm->locked_vm += vwork->npage;
+       up_write(&mm->mmap_sem);
+       mmput(mm);
+       kfree(vwork);
+}
+
+static void vfio_lock_acct(long npage)
+{
+       struct vwork *vwork;
+       struct mm_struct *mm;
+
+       if (!current->mm)
+               return; /* process exited */
+
+       if (down_write_trylock(&current->mm->mmap_sem)) {
+               current->mm->locked_vm += npage;
+               up_write(&current->mm->mmap_sem);
+               return;
+       }
+
+       /*
+        * Couldn't get mmap_sem lock, so must setup to update
+        * mm->locked_vm later. If locked_vm were atomic, we
+        * wouldn't need this silliness
+        */
+       vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
+       if (!vwork)
+               return;
+       mm = get_task_mm(current);
+       if (!mm) {
+               kfree(vwork);
+               return;
+       }
+       INIT_WORK(&vwork->work, vfio_lock_acct_bg);
+       vwork->mm = mm;
+       vwork->npage = npage;
+       schedule_work(&vwork->work);
+}
+
+/*
+ * Some mappings aren't backed by a struct page, for example an mmap'd
+ * MMIO range for our own or another device.  These use a different
+ * pfn conversion and shouldn't be tracked as locked pages.
+ */
+static bool is_invalid_reserved_pfn(unsigned long pfn)
+{
+       if (pfn_valid(pfn)) {
+               bool reserved;
+               struct page *tail = pfn_to_page(pfn);
+               struct page *head = compound_trans_head(tail);
+               reserved = !!(PageReserved(head));
+               if (head != tail) {
+                       /*
+                        * "head" is not a dangling pointer
+                        * (compound_trans_head takes care of that)
+                        * but the hugepage may have been split
+                        * from under us (and we may not hold a
+                        * reference count on the head page so it can
+                        * be reused before we run PageReferenced), so
+                        * we've to check PageTail before returning
+                        * what we just read.
+                        */
+                       smp_rmb();
+                       if (PageTail(tail))
+                               return reserved;
+               }
+               return PageReserved(tail);
+       }
+
+       return true;
+}
+
+static int put_pfn(unsigned long pfn, int prot)
+{
+       if (!is_invalid_reserved_pfn(pfn)) {
+               struct page *page = pfn_to_page(pfn);
+               if (prot & IOMMU_WRITE)
+                       SetPageDirty(page);
+               put_page(page);
+               return 1;
+       }
+       return 0;
+}
+
+/* Unmap DMA region */
+static long __vfio_dma_do_unmap(struct vfio_iommu *iommu, dma_addr_t iova,
+                            long npage, int prot)
+{
+       long i, unlocked = 0;
+
+       for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
+               unsigned long pfn;
+
+               pfn = iommu_iova_to_phys(iommu->domain, iova) >> PAGE_SHIFT;
+               if (pfn) {
+                       iommu_unmap(iommu->domain, iova, PAGE_SIZE);
+                       unlocked += put_pfn(pfn, prot);
+               }
+       }
+       return unlocked;
+}
+
+static void vfio_dma_unmap(struct vfio_iommu *iommu, dma_addr_t iova,
+                          long npage, int prot)
+{
+       long unlocked;
+
+       unlocked = __vfio_dma_do_unmap(iommu, iova, npage, prot);
+       vfio_lock_acct(-unlocked);
+}
+
+static int vaddr_get_pfn(unsigned long vaddr, int prot, unsigned long *pfn)
+{
+       struct page *page[1];
+       struct vm_area_struct *vma;
+       int ret = -EFAULT;
+
+       if (get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), page) == 1) {
+               *pfn = page_to_pfn(page[0]);
+               return 0;
+       }
+
+       down_read(&current->mm->mmap_sem);
+
+       vma = find_vma_intersection(current->mm, vaddr, vaddr + 1);
+
+       if (vma && vma->vm_flags & VM_PFNMAP) {
+               *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
+               if (is_invalid_reserved_pfn(*pfn))
+                       ret = 0;
+       }
+
+       up_read(&current->mm->mmap_sem);
+
+       return ret;
+}
+
+/* Map DMA region */
+static int __vfio_dma_map(struct vfio_iommu *iommu, dma_addr_t iova,
+                         unsigned long vaddr, long npage, int prot)
+{
+       dma_addr_t start = iova;
+       long i, locked = 0;
+       int ret;
+
+       /* Verify that pages are not already mapped */
+       for (i = 0; i < npage; i++, iova += PAGE_SIZE)
+               if (iommu_iova_to_phys(iommu->domain, iova))
+                       return -EBUSY;
+
+       iova = start;
+
+       if (iommu->cache)
+               prot |= IOMMU_CACHE;
+
+       /*
+        * XXX We break mappings into pages and use get_user_pages_fast to
+        * pin the pages in memory.  It's been suggested that mlock might
+        * provide a more efficient mechanism, but nothing prevents the
+        * user from munlocking the pages, which could then allow the user
+        * access to random host memory.  We also have no guarantee from the
+        * IOMMU API that the iommu driver can unmap sub-pages of previous
+        * mappings.  This means we might lose an entire range if a single
+        * page within it is unmapped.  Single page mappings are inefficient,
+        * but provide the most flexibility for now.
+        */
+       for (i = 0; i < npage; i++, iova += PAGE_SIZE, vaddr += PAGE_SIZE) {
+               unsigned long pfn = 0;
+
+               ret = vaddr_get_pfn(vaddr, prot, &pfn);
+               if (ret) {
+                       __vfio_dma_do_unmap(iommu, start, i, prot);
+                       return ret;
+               }
+
+               /*
+                * Only add actual locked pages to accounting
+                * XXX We're effectively marking a page locked for every
+                * IOVA page even though it's possible the user could be
+                * backing multiple IOVAs with the same vaddr.  This over-
+                * penalizes the user process, but we currently have no
+                * easy way to do this properly.
+                */
+               if (!is_invalid_reserved_pfn(pfn))
+                       locked++;
+
+               ret = iommu_map(iommu->domain, iova,
+                               (phys_addr_t)pfn << PAGE_SHIFT,
+                               PAGE_SIZE, prot);
+               if (ret) {
+                       /* Back out mappings on error */
+                       put_pfn(pfn, prot);
+                       __vfio_dma_do_unmap(iommu, start, i, prot);
+                       return ret;
+               }
+       }
+       vfio_lock_acct(locked);
+       return 0;
+}
+
+static inline bool ranges_overlap(dma_addr_t start1, size_t size1,
+                                 dma_addr_t start2, size_t size2)
+{
+       if (start1 < start2)
+               return (start2 - start1 < size1);
+       else if (start2 < start1)
+               return (start1 - start2 < size2);
+       return (size1 > 0 && size2 > 0);
+}
+
+static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
+                                               dma_addr_t start, size_t size)
+{
+       struct vfio_dma *dma;
+
+       list_for_each_entry(dma, &iommu->dma_list, next) {
+               if (ranges_overlap(dma->iova, NPAGE_TO_SIZE(dma->npage),
+                                  start, size))
+                       return dma;
+       }
+       return NULL;
+}
+
+static long vfio_remove_dma_overlap(struct vfio_iommu *iommu, dma_addr_t start,
+                                   size_t size, struct vfio_dma *dma)
+{
+       struct vfio_dma *split;
+       long npage_lo, npage_hi;
+
+       /* Existing dma region is completely covered, unmap all */
+       if (start <= dma->iova &&
+           start + size >= dma->iova + NPAGE_TO_SIZE(dma->npage)) {
+               vfio_dma_unmap(iommu, dma->iova, dma->npage, dma->prot);
+               list_del(&dma->next);
+               npage_lo = dma->npage;
+               kfree(dma);
+               return npage_lo;
+       }
+
+       /* Overlap low address of existing range */
+       if (start <= dma->iova) {
+               size_t overlap;
+
+               overlap = start + size - dma->iova;
+               npage_lo = overlap >> PAGE_SHIFT;
+
+               vfio_dma_unmap(iommu, dma->iova, npage_lo, dma->prot);
+               dma->iova += overlap;
+               dma->vaddr += overlap;
+               dma->npage -= npage_lo;
+               return npage_lo;
+       }
+
+       /* Overlap high address of existing range */
+       if (start + size >= dma->iova + NPAGE_TO_SIZE(dma->npage)) {
+               size_t overlap;
+
+               overlap = dma->iova + NPAGE_TO_SIZE(dma->npage) - start;
+               npage_hi = overlap >> PAGE_SHIFT;
+
+               vfio_dma_unmap(iommu, start, npage_hi, dma->prot);
+               dma->npage -= npage_hi;
+               return npage_hi;
+       }
+
+       /* Split existing */
+       npage_lo = (start - dma->iova) >> PAGE_SHIFT;
+       npage_hi = dma->npage - (size >> PAGE_SHIFT) - npage_lo;
+
+       split = kzalloc(sizeof *split, GFP_KERNEL);
+       if (!split)
+               return -ENOMEM;
+
+       vfio_dma_unmap(iommu, start, size >> PAGE_SHIFT, dma->prot);
+
+       dma->npage = npage_lo;
+
+       split->npage = npage_hi;
+       split->iova = start + size;
+       split->vaddr = dma->vaddr + NPAGE_TO_SIZE(npage_lo) + size;
+       split->prot = dma->prot;
+       list_add(&split->next, &iommu->dma_list);
+       return size >> PAGE_SHIFT;
+}
+
+static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
+                            struct vfio_iommu_type1_dma_unmap *unmap)
+{
+       long ret = 0, npage = unmap->size >> PAGE_SHIFT;
+       struct vfio_dma *dma, *tmp;
+       uint64_t mask;
+
+       mask = ((uint64_t)1 << __ffs(iommu->domain->ops->pgsize_bitmap)) - 1;
+
+       if (unmap->iova & mask)
+               return -EINVAL;
+       if (unmap->size & mask)
+               return -EINVAL;
+
+       /* XXX We still break these down into PAGE_SIZE */
+       WARN_ON(mask & PAGE_MASK);
+
+       mutex_lock(&iommu->lock);
+
+       list_for_each_entry_safe(dma, tmp, &iommu->dma_list, next) {
+               if (ranges_overlap(dma->iova, NPAGE_TO_SIZE(dma->npage),
+                                  unmap->iova, unmap->size)) {
+                       ret = vfio_remove_dma_overlap(iommu, unmap->iova,
+                                                     unmap->size, dma);
+                       if (ret > 0)
+                               npage -= ret;
+                       if (ret < 0 || npage == 0)
+                               break;
+               }
+       }
+       mutex_unlock(&iommu->lock);
+       return ret > 0 ? 0 : (int)ret;
+}
+
+static int vfio_dma_do_map(struct vfio_iommu *iommu,
+                          struct vfio_iommu_type1_dma_map *map)
+{
+       struct vfio_dma *dma, *pdma = NULL;
+       dma_addr_t iova = map->iova;
+       unsigned long locked, lock_limit, vaddr = map->vaddr;
+       size_t size = map->size;
+       int ret = 0, prot = 0;
+       uint64_t mask;
+       long npage;
+
+       mask = ((uint64_t)1 << __ffs(iommu->domain->ops->pgsize_bitmap)) - 1;
+
+       /* READ/WRITE from device perspective */
+       if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
+               prot |= IOMMU_WRITE;
+       if (map->flags & VFIO_DMA_MAP_FLAG_READ)
+               prot |= IOMMU_READ;
+
+       if (!prot)
+               return -EINVAL; /* No READ/WRITE? */
+
+       if (vaddr & mask)
+               return -EINVAL;
+       if (iova & mask)
+               return -EINVAL;
+       if (size & mask)
+               return -EINVAL;
+
+       /* XXX We still break these down into PAGE_SIZE */
+       WARN_ON(mask & PAGE_MASK);
+
+       /* Don't allow IOVA wrap */
+       if (iova + size && iova + size < iova)
+               return -EINVAL;
+
+       /* Don't allow virtual address wrap */
+       if (vaddr + size && vaddr + size < vaddr)
+               return -EINVAL;
+
+       npage = size >> PAGE_SHIFT;
+       if (!npage)
+               return -EINVAL;
+
+       mutex_lock(&iommu->lock);
+
+       if (vfio_find_dma(iommu, iova, size)) {
+               ret = -EBUSY;
+               goto out_lock;
+       }
+
+       /* account for locked pages */
+       locked = current->mm->locked_vm + npage;
+       lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+       if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
+               pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
+                       __func__, rlimit(RLIMIT_MEMLOCK));
+               ret = -ENOMEM;
+               goto out_lock;
+       }
+
+       ret = __vfio_dma_map(iommu, iova, vaddr, npage, prot);
+       if (ret)
+               goto out_lock;
+
+       /* Check if we abut a region below - nothing below 0 */
+       if (iova) {
+               dma = vfio_find_dma(iommu, iova - 1, 1);
+               if (dma && dma->prot == prot &&
+                   dma->vaddr + NPAGE_TO_SIZE(dma->npage) == vaddr) {
+
+                       dma->npage += npage;
+                       iova = dma->iova;
+                       vaddr = dma->vaddr;
+                       npage = dma->npage;
+                       size = NPAGE_TO_SIZE(npage);
+
+                       pdma = dma;
+               }
+       }
+
+       /* Check if we abut a region above - nothing above ~0 + 1 */
+       if (iova + size) {
+               dma = vfio_find_dma(iommu, iova + size, 1);
+               if (dma && dma->prot == prot &&
+                   dma->vaddr == vaddr + size) {
+
+                       dma->npage += npage;
+                       dma->iova = iova;
+                       dma->vaddr = vaddr;
+
+                       /*
+                        * If merged above and below, remove previously
+                        * merged entry.  New entry covers it.
+                        */
+                       if (pdma) {
+                               list_del(&pdma->next);
+                               kfree(pdma);
+                       }
+                       pdma = dma;
+               }
+       }
+
+       /* Isolated, new region */
+       if (!pdma) {
+               dma = kzalloc(sizeof *dma, GFP_KERNEL);
+               if (!dma) {
+                       ret = -ENOMEM;
+                       vfio_dma_unmap(iommu, iova, npage, prot);
+                       goto out_lock;
+               }
+
+               dma->npage = npage;
+               dma->iova = iova;
+               dma->vaddr = vaddr;
+               dma->prot = prot;
+               list_add(&dma->next, &iommu->dma_list);
+       }
+
+out_lock:
+       mutex_unlock(&iommu->lock);
+       return ret;
+}
+
+static int vfio_iommu_type1_attach_group(void *iommu_data,
+                                        struct iommu_group *iommu_group)
+{
+       struct vfio_iommu *iommu = iommu_data;
+       struct vfio_group *group, *tmp;
+       int ret;
+
+       group = kzalloc(sizeof(*group), GFP_KERNEL);
+       if (!group)
+               return -ENOMEM;
+
+       mutex_lock(&iommu->lock);
+
+       list_for_each_entry(tmp, &iommu->group_list, next) {
+               if (tmp->iommu_group == iommu_group) {
+                       mutex_unlock(&iommu->lock);
+                       kfree(group);
+                       return -EINVAL;
+               }
+       }
+
+       /*
+        * TODO: Domain have capabilities that might change as we add
+        * groups (see iommu->cache, currently never set).  Check for
+        * them and potentially disallow groups to be attached when it
+        * would change capabilities (ugh).
+        */
+       ret = iommu_attach_group(iommu->domain, iommu_group);
+       if (ret) {
+               mutex_unlock(&iommu->lock);
+               kfree(group);
+               return ret;
+       }
+
+       group->iommu_group = iommu_group;
+       list_add(&group->next, &iommu->group_list);
+
+       mutex_unlock(&iommu->lock);
+
+       return 0;
+}
+
+static void vfio_iommu_type1_detach_group(void *iommu_data,
+                                         struct iommu_group *iommu_group)
+{
+       struct vfio_iommu *iommu = iommu_data;
+       struct vfio_group *group;
+
+       mutex_lock(&iommu->lock);
+
+       list_for_each_entry(group, &iommu->group_list, next) {
+               if (group->iommu_group == iommu_group) {
+                       iommu_detach_group(iommu->domain, iommu_group);
+                       list_del(&group->next);
+                       kfree(group);
+                       break;
+               }
+       }
+
+       mutex_unlock(&iommu->lock);
+}
+
+static void *vfio_iommu_type1_open(unsigned long arg)
+{
+       struct vfio_iommu *iommu;
+
+       if (arg != VFIO_TYPE1_IOMMU)
+               return ERR_PTR(-EINVAL);
+
+       iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
+       if (!iommu)
+               return ERR_PTR(-ENOMEM);
+
+       INIT_LIST_HEAD(&iommu->group_list);
+       INIT_LIST_HEAD(&iommu->dma_list);
+       mutex_init(&iommu->lock);
+
+       /*
+        * Wish we didn't have to know about bus_type here.
+        */
+       iommu->domain = iommu_domain_alloc(&pci_bus_type);
+       if (!iommu->domain) {
+               kfree(iommu);
+               return ERR_PTR(-EIO);
+       }
+
+       /*
+        * Wish we could specify required capabilities rather than create
+        * a domain, see what comes out and hope it doesn't change along
+        * the way.  Fortunately we know interrupt remapping is global for
+        * our iommus.
+        */
+       if (!allow_unsafe_interrupts &&
+           !iommu_domain_has_cap(iommu->domain, IOMMU_CAP_INTR_REMAP)) {
+               pr_warn("%s: No interrupt remapping support.  Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
+                      __func__);
+               iommu_domain_free(iommu->domain);
+               kfree(iommu);
+               return ERR_PTR(-EPERM);
+       }
+
+       return iommu;
+}
+
+static void vfio_iommu_type1_release(void *iommu_data)
+{
+       struct vfio_iommu *iommu = iommu_data;
+       struct vfio_group *group, *group_tmp;
+       struct vfio_dma *dma, *dma_tmp;
+
+       list_for_each_entry_safe(group, group_tmp, &iommu->group_list, next) {
+               iommu_detach_group(iommu->domain, group->iommu_group);
+               list_del(&group->next);
+               kfree(group);
+       }
+
+       list_for_each_entry_safe(dma, dma_tmp, &iommu->dma_list, next) {
+               vfio_dma_unmap(iommu, dma->iova, dma->npage, dma->prot);
+               list_del(&dma->next);
+               kfree(dma);
+       }
+
+       iommu_domain_free(iommu->domain);
+       iommu->domain = NULL;
+       kfree(iommu);
+}
+
+static long vfio_iommu_type1_ioctl(void *iommu_data,
+                                  unsigned int cmd, unsigned long arg)
+{
+       struct vfio_iommu *iommu = iommu_data;
+       unsigned long minsz;
+
+       if (cmd == VFIO_CHECK_EXTENSION) {
+               switch (arg) {
+               case VFIO_TYPE1_IOMMU:
+                       return 1;
+               default:
+                       return 0;
+               }
+       } else if (cmd == VFIO_IOMMU_GET_INFO) {
+               struct vfio_iommu_type1_info info;
+
+               minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
+
+               if (copy_from_user(&info, (void __user *)arg, minsz))
+                       return -EFAULT;
+
+               if (info.argsz < minsz)
+                       return -EINVAL;
+
+               info.flags = 0;
+
+               info.iova_pgsizes = iommu->domain->ops->pgsize_bitmap;
+
+               return copy_to_user((void __user *)arg, &info, minsz);
+
+       } else if (cmd == VFIO_IOMMU_MAP_DMA) {
+               struct vfio_iommu_type1_dma_map map;
+               uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
+                               VFIO_DMA_MAP_FLAG_WRITE;
+
+               minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
+
+               if (copy_from_user(&map, (void __user *)arg, minsz))
+                       return -EFAULT;
+
+               if (map.argsz < minsz || map.flags & ~mask)
+                       return -EINVAL;
+
+               return vfio_dma_do_map(iommu, &map);
+
+       } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
+               struct vfio_iommu_type1_dma_unmap unmap;
+
+               minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
+
+               if (copy_from_user(&unmap, (void __user *)arg, minsz))
+                       return -EFAULT;
+
+               if (unmap.argsz < minsz || unmap.flags)
+                       return -EINVAL;
+
+               return vfio_dma_do_unmap(iommu, &unmap);
+       }
+
+       return -ENOTTY;
+}
+
+static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
+       .name           = "vfio-iommu-type1",
+       .owner          = THIS_MODULE,
+       .open           = vfio_iommu_type1_open,
+       .release        = vfio_iommu_type1_release,
+       .ioctl          = vfio_iommu_type1_ioctl,
+       .attach_group   = vfio_iommu_type1_attach_group,
+       .detach_group   = vfio_iommu_type1_detach_group,
+};
+
+static int __init vfio_iommu_type1_init(void)
+{
+       if (!iommu_present(&pci_bus_type))
+               return -ENODEV;
+
+       return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
+}
+
+static void __exit vfio_iommu_type1_cleanup(void)
+{
+       vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
+}
+
+module_init(vfio_iommu_type1_init);
+module_exit(vfio_iommu_type1_cleanup);
+
+MODULE_VERSION(DRIVER_VERSION);
+MODULE_LICENSE("GPL v2");
+MODULE_AUTHOR(DRIVER_AUTHOR);
+MODULE_DESCRIPTION(DRIVER_DESC);
index 03e56a5154b6f6bcf6c3d333d4ff7612b2a33be4..acb046fd5b70bfa8aa715197bf27304916786ff9 100644 (file)
@@ -98,7 +98,7 @@ extern void vfio_unregister_iommu_driver(
 
 /* Extensions */
 
-/* None yet */
+#define VFIO_TYPE1_IOMMU               1
 
 /*
  * The IOCTL interface is designed for extensibility by embedding the
@@ -364,4 +364,56 @@ struct vfio_irq_set {
  */
 #define VFIO_DEVICE_RESET              _IO(VFIO_TYPE, VFIO_BASE + 11)
 
+/* -------- API for Type1 VFIO IOMMU -------- */
+
+/**
+ * VFIO_IOMMU_GET_INFO - _IOR(VFIO_TYPE, VFIO_BASE + 12, struct vfio_iommu_info)
+ *
+ * Retrieve information about the IOMMU object. Fills in provided
+ * struct vfio_iommu_info. Caller sets argsz.
+ *
+ * XXX Should we do these by CHECK_EXTENSION too?
+ */
+struct vfio_iommu_type1_info {
+       __u32   argsz;
+       __u32   flags;
+#define VFIO_IOMMU_INFO_PGSIZES (1 << 0)       /* supported page sizes info */
+       __u64   iova_pgsizes;           /* Bitmap of supported page sizes */
+};
+
+#define VFIO_IOMMU_GET_INFO _IO(VFIO_TYPE, VFIO_BASE + 12)
+
+/**
+ * VFIO_IOMMU_MAP_DMA - _IOW(VFIO_TYPE, VFIO_BASE + 13, struct vfio_dma_map)
+ *
+ * Map process virtual addresses to IO virtual addresses using the
+ * provided struct vfio_dma_map. Caller sets argsz. READ &/ WRITE required.
+ */
+struct vfio_iommu_type1_dma_map {
+       __u32   argsz;
+       __u32   flags;
+#define VFIO_DMA_MAP_FLAG_READ (1 << 0)                /* readable from device */
+#define VFIO_DMA_MAP_FLAG_WRITE (1 << 1)       /* writable from device */
+       __u64   vaddr;                          /* Process virtual address */
+       __u64   iova;                           /* IO virtual address */
+       __u64   size;                           /* Size of mapping (bytes) */
+};
+
+#define VFIO_IOMMU_MAP_DMA _IO(VFIO_TYPE, VFIO_BASE + 13)
+
+/**
+ * VFIO_IOMMU_UNMAP_DMA - _IOW(VFIO_TYPE, VFIO_BASE + 14, struct vfio_dma_unmap)
+ *
+ * Unmap IO virtual addresses using the provided struct vfio_dma_unmap.
+ * Caller sets argsz.
+ */
+struct vfio_iommu_type1_dma_unmap {
+       __u32   argsz;
+       __u32   flags;
+       __u64   iova;                           /* IO virtual address */
+       __u64   size;                           /* Size of mapping (bytes) */
+};
+
+#define VFIO_IOMMU_UNMAP_DMA _IO(VFIO_TYPE, VFIO_BASE + 14)
+
 #endif /* VFIO_H */