virt/kvm/iommu.c

   1 /*
   2  * Copyright (c) 2006, Intel Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Copyright (C) 2006-2008 Intel Corporation
  18  * Copyright IBM Corporation, 2008
  19  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
  20  *
  21  * Author: Allen M. Kay <allen.m.kay@intel.com>
  22  * Author: Weidong Han <weidong.han@intel.com>
  23  * Author: Ben-Ami Yassour <benami@il.ibm.com>
  24  */
  25
  26 #include <linux/list.h>
  27 #include <linux/kvm_host.h>
  28 #include <linux/module.h>
  29 #include <linux/pci.h>
  30 #include <linux/stat.h>
  31 #include <linux/dmar.h>
  32 #include <linux/iommu.h>
  33 #include <linux/intel-iommu.h>
  34
  35 static bool allow_unsafe_assigned_interrupts;
  36 module_param_named(allow_unsafe_assigned_interrupts,
  37                    allow_unsafe_assigned_interrupts, bool, S_IRUGO | S_IWUSR);
  38 MODULE_PARM_DESC(allow_unsafe_assigned_interrupts,
  39  "Enable device assignment on platforms without interrupt remapping support.");
  40
  41 static int kvm_iommu_unmap_memslots(struct kvm *kvm);
  42 static void kvm_iommu_put_pages(struct kvm *kvm,
  43                                 gfn_t base_gfn, unsigned long npages);
  44
  45 static pfn_t kvm_pin_pages(struct kvm_memory_slot *slot, gfn_t gfn,
  46                            unsigned long npages)
  47 {
  48         gfn_t end_gfn;
  49         pfn_t pfn;
  50
  51         pfn     = gfn_to_pfn_memslot(slot, gfn);
  52         end_gfn = gfn + npages;
  53         gfn    += 1;
  54
  55         if (is_error_noslot_pfn(pfn))
  56                 return pfn;
  57
  58         while (gfn < end_gfn)
  59                 gfn_to_pfn_memslot(slot, gfn++);
  60
  61         return pfn;
  62 }
  63
  64 static void kvm_unpin_pages(struct kvm *kvm, pfn_t pfn, unsigned long npages)
  65 {
  66         unsigned long i;
  67
  68         for (i = 0; i < npages; ++i)
  69                 kvm_release_pfn_clean(pfn + i);
  70 }
  71
  72 int kvm_iommu_map_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
  73 {
  74         gfn_t gfn, end_gfn;
  75         pfn_t pfn;
  76         int r = 0;
  77         struct iommu_domain *domain = kvm->arch.iommu_domain;
  78         int flags;
  79
  80         /* check if iommu exists and in use */
  81         if (!domain)
  82                 return 0;
  83
  84         gfn     = slot->base_gfn;
  85         end_gfn = gfn + slot->npages;
  86
  87         flags = IOMMU_READ;
  88         if (!(slot->flags & KVM_MEM_READONLY))
  89                 flags |= IOMMU_WRITE;
  90         if (!kvm->arch.iommu_noncoherent)
  91                 flags |= IOMMU_CACHE;
  92
  93
  94         while (gfn < end_gfn) {
  95                 unsigned long page_size;
  96
  97                 /* Check if already mapped */
  98                 if (iommu_iova_to_phys(domain, gfn_to_gpa(gfn))) {
  99                         gfn += 1;
 100                         continue;
 101                 }
 102
 103                 /* Get the page size we could use to map */
 104                 page_size = kvm_host_page_size(kvm, gfn);
 105
 106                 /* Make sure the page_size does not exceed the memslot */
 107                 while ((gfn + (page_size >> PAGE_SHIFT)) > end_gfn)
 108                         page_size >>= 1;
 109
 110                 /* Make sure gfn is aligned to the page size we want to map */
 111                 while ((gfn << PAGE_SHIFT) & (page_size - 1))
 112                         page_size >>= 1;
 113
 114                 /* Make sure hva is aligned to the page size we want to map */
 115                 while (__gfn_to_hva_memslot(slot, gfn) & (page_size - 1))
 116                         page_size >>= 1;
 117
 118                 /*
 119                  * Pin all pages we are about to map in memory. This is
 120                  * important because we unmap and unpin in 4kb steps later.
 121                  */
 122                 pfn = kvm_pin_pages(slot, gfn, page_size >> PAGE_SHIFT);
 123                 if (is_error_noslot_pfn(pfn)) {
 124                         gfn += 1;
 125                         continue;
 126                 }
 127
 128                 /* Map into IO address space */
 129                 r = iommu_map(domain, gfn_to_gpa(gfn), pfn_to_hpa(pfn),
 130                               page_size, flags);
 131                 if (r) {
 132                         printk(KERN_ERR "kvm_iommu_map_address:"
 133                                "iommu failed to map pfn=%llx\n", pfn);
 134                         kvm_unpin_pages(kvm, pfn, page_size >> PAGE_SHIFT);
 135                         goto unmap_pages;
 136                 }
 137
 138                 gfn += page_size >> PAGE_SHIFT;
 139
 140
 141         }
 142
 143         return 0;
 144
 145 unmap_pages:
 146         kvm_iommu_put_pages(kvm, slot->base_gfn, gfn - slot->base_gfn);
 147         return r;
 148 }
 149
 150 static int kvm_iommu_map_memslots(struct kvm *kvm)
 151 {
 152         int idx, r = 0;
 153         struct kvm_memslots *slots;
 154         struct kvm_memory_slot *memslot;
 155
 156         if (kvm->arch.iommu_noncoherent)
 157                 kvm_arch_register_noncoherent_dma(kvm);
 158
 159         idx = srcu_read_lock(&kvm->srcu);
 160         slots = kvm_memslots(kvm);
 161
 162         kvm_for_each_memslot(memslot, slots) {
 163                 r = kvm_iommu_map_pages(kvm, memslot);
 164                 if (r)
 165                         break;
 166         }
 167         srcu_read_unlock(&kvm->srcu, idx);
 168
 169         return r;
 170 }
 171
 172 int kvm_assign_device(struct kvm *kvm,
 173                       struct kvm_assigned_dev_kernel *assigned_dev)
 174 {
 175         struct pci_dev *pdev = NULL;
 176         struct iommu_domain *domain = kvm->arch.iommu_domain;
 177         int r;
 178         bool noncoherent;
 179
 180         /* check if iommu exists and in use */
 181         if (!domain)
 182                 return 0;
 183
 184         pdev = assigned_dev->dev;
 185         if (pdev == NULL)
 186                 return -ENODEV;
 187
 188         r = iommu_attach_device(domain, &pdev->dev);
 189         if (r) {
 190                 dev_err(&pdev->dev, "kvm assign device failed ret %d", r);
 191                 return r;
 192         }
 193
 194         noncoherent = !iommu_capable(&pci_bus_type, IOMMU_CAP_CACHE_COHERENCY);
 195
 196         /* Check if need to update IOMMU page table for guest memory */
 197         if (noncoherent != kvm->arch.iommu_noncoherent) {
 198                 kvm_iommu_unmap_memslots(kvm);
 199                 kvm->arch.iommu_noncoherent = noncoherent;
 200                 r = kvm_iommu_map_memslots(kvm);
 201                 if (r)
 202                         goto out_unmap;
 203         }
 204
 205         pci_set_dev_assigned(pdev);
 206
 207         dev_info(&pdev->dev, "kvm assign device\n");
 208
 209         return 0;
 210 out_unmap:
 211         kvm_iommu_unmap_memslots(kvm);
 212         return r;
 213 }
 214
 215 int kvm_deassign_device(struct kvm *kvm,
 216                         struct kvm_assigned_dev_kernel *assigned_dev)
 217 {
 218         struct iommu_domain *domain = kvm->arch.iommu_domain;
 219         struct pci_dev *pdev = NULL;
 220
 221         /* check if iommu exists and in use */
 222         if (!domain)
 223                 return 0;
 224
 225         pdev = assigned_dev->dev;
 226         if (pdev == NULL)
 227                 return -ENODEV;
 228
 229         iommu_detach_device(domain, &pdev->dev);
 230
 231         pci_clear_dev_assigned(pdev);
 232
 233         dev_info(&pdev->dev, "kvm deassign device\n");
 234
 235         return 0;
 236 }
 237
 238 int kvm_iommu_map_guest(struct kvm *kvm)
 239 {
 240         int r;
 241
 242         if (!iommu_present(&pci_bus_type)) {
 243                 printk(KERN_ERR "%s: iommu not found\n", __func__);
 244                 return -ENODEV;
 245         }
 246
 247         mutex_lock(&kvm->slots_lock);
 248
 249         kvm->arch.iommu_domain = iommu_domain_alloc(&pci_bus_type);
 250         if (!kvm->arch.iommu_domain) {
 251                 r = -ENOMEM;
 252                 goto out_unlock;
 253         }
 254
 255         if (!allow_unsafe_assigned_interrupts &&
 256             !iommu_capable(&pci_bus_type, IOMMU_CAP_INTR_REMAP)) {
 257                 printk(KERN_WARNING "%s: No interrupt remapping support,"
 258                        " disallowing device assignment."
 259                        " Re-enble with \"allow_unsafe_assigned_interrupts=1\""
 260                        " module option.\n", __func__);
 261                 iommu_domain_free(kvm->arch.iommu_domain);
 262                 kvm->arch.iommu_domain = NULL;
 263                 r = -EPERM;
 264                 goto out_unlock;
 265         }
 266
 267         r = kvm_iommu_map_memslots(kvm);
 268         if (r)
 269                 kvm_iommu_unmap_memslots(kvm);
 270
 271 out_unlock:
 272         mutex_unlock(&kvm->slots_lock);
 273         return r;
 274 }
 275
 276 static void kvm_iommu_put_pages(struct kvm *kvm,
 277                                 gfn_t base_gfn, unsigned long npages)
 278 {
 279         struct iommu_domain *domain;
 280         gfn_t end_gfn, gfn;
 281         pfn_t pfn;
 282         u64 phys;
 283
 284         domain  = kvm->arch.iommu_domain;
 285         end_gfn = base_gfn + npages;
 286         gfn     = base_gfn;
 287
 288         /* check if iommu exists and in use */
 289         if (!domain)
 290                 return;
 291
 292         while (gfn < end_gfn) {
 293                 unsigned long unmap_pages;
 294                 size_t size;
 295
 296                 /* Get physical address */
 297                 phys = iommu_iova_to_phys(domain, gfn_to_gpa(gfn));
 298
 299                 if (!phys) {
 300                         gfn++;
 301                         continue;
 302                 }
 303
 304                 pfn  = phys >> PAGE_SHIFT;
 305
 306                 /* Unmap address from IO address space */
 307                 size       = iommu_unmap(domain, gfn_to_gpa(gfn), PAGE_SIZE);
 308                 unmap_pages = 1ULL << get_order(size);
 309
 310                 /* Unpin all pages we just unmapped to not leak any memory */
 311                 kvm_unpin_pages(kvm, pfn, unmap_pages);
 312
 313                 gfn += unmap_pages;
 314         }
 315 }
 316
 317 void kvm_iommu_unmap_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
 318 {
 319         kvm_iommu_put_pages(kvm, slot->base_gfn, slot->npages);
 320 }
 321
 322 static int kvm_iommu_unmap_memslots(struct kvm *kvm)
 323 {
 324         int idx;
 325         struct kvm_memslots *slots;
 326         struct kvm_memory_slot *memslot;
 327
 328         idx = srcu_read_lock(&kvm->srcu);
 329         slots = kvm_memslots(kvm);
 330
 331         kvm_for_each_memslot(memslot, slots)
 332                 kvm_iommu_unmap_pages(kvm, memslot);
 333
 334         srcu_read_unlock(&kvm->srcu, idx);
 335
 336         if (kvm->arch.iommu_noncoherent)
 337                 kvm_arch_unregister_noncoherent_dma(kvm);
 338
 339         return 0;
 340 }
 341
 342 int kvm_iommu_unmap_guest(struct kvm *kvm)
 343 {
 344         struct iommu_domain *domain = kvm->arch.iommu_domain;
 345
 346         /* check if iommu exists and in use */
 347         if (!domain)
 348                 return 0;
 349
 350         mutex_lock(&kvm->slots_lock);
 351         kvm_iommu_unmap_memslots(kvm);
 352         kvm->arch.iommu_domain = NULL;
 353         kvm->arch.iommu_noncoherent = false;
 354         mutex_unlock(&kvm->slots_lock);
 355
 356         iommu_domain_free(domain);
 357         return 0;
 358 }