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 size)
  47 {
  48         gfn_t end_gfn;
  49         pfn_t pfn;
  50
  51         pfn     = gfn_to_pfn_memslot(slot, gfn);
  52         end_gfn = gfn + (size >> PAGE_SHIFT);
  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 int kvm_iommu_map_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
  65 {
  66         gfn_t gfn, end_gfn;
  67         pfn_t pfn;
  68         int r = 0;
  69         struct iommu_domain *domain = kvm->arch.iommu_domain;
  70         int flags;
  71
  72         /* check if iommu exists and in use */
  73         if (!domain)
  74                 return 0;
  75
  76         gfn     = slot->base_gfn;
  77         end_gfn = gfn + slot->npages;
  78
  79         flags = IOMMU_READ;
  80         if (!(slot->flags & KVM_MEM_READONLY))
  81                 flags |= IOMMU_WRITE;
  82         if (!kvm->arch.iommu_noncoherent)
  83                 flags |= IOMMU_CACHE;
  84
  85
  86         while (gfn < end_gfn) {
  87                 unsigned long page_size;
  88
  89                 /* Check if already mapped */
  90                 if (iommu_iova_to_phys(domain, gfn_to_gpa(gfn))) {
  91                         gfn += 1;
  92                         continue;
  93                 }
  94
  95                 /* Get the page size we could use to map */
  96                 page_size = kvm_host_page_size(kvm, gfn);
  97
  98                 /* Make sure the page_size does not exceed the memslot */
  99                 while ((gfn + (page_size >> PAGE_SHIFT)) > end_gfn)
 100                         page_size >>= 1;
 101
 102                 /* Make sure gfn is aligned to the page size we want to map */
 103                 while ((gfn << PAGE_SHIFT) & (page_size - 1))
 104                         page_size >>= 1;
 105
 106                 /*
 107                  * Pin all pages we are about to map in memory. This is
 108                  * important because we unmap and unpin in 4kb steps later.
 109                  */
 110                 pfn = kvm_pin_pages(slot, gfn, page_size);
 111                 if (is_error_noslot_pfn(pfn)) {
 112                         gfn += 1;
 113                         continue;
 114                 }
 115
 116                 /* Map into IO address space */
 117                 r = iommu_map(domain, gfn_to_gpa(gfn), pfn_to_hpa(pfn),
 118                               page_size, flags);
 119                 if (r) {
 120                         printk(KERN_ERR "kvm_iommu_map_address:"
 121                                "iommu failed to map pfn=%llx\n", pfn);
 122                         goto unmap_pages;
 123                 }
 124
 125                 gfn += page_size >> PAGE_SHIFT;
 126
 127
 128         }
 129
 130         return 0;
 131
 132 unmap_pages:
 133         kvm_iommu_put_pages(kvm, slot->base_gfn, gfn);
 134         return r;
 135 }
 136
 137 static int kvm_iommu_map_memslots(struct kvm *kvm)
 138 {
 139         int idx, r = 0;
 140         struct kvm_memslots *slots;
 141         struct kvm_memory_slot *memslot;
 142
 143         idx = srcu_read_lock(&kvm->srcu);
 144         slots = kvm_memslots(kvm);
 145
 146         kvm_for_each_memslot(memslot, slots) {
 147                 r = kvm_iommu_map_pages(kvm, memslot);
 148                 if (r)
 149                         break;
 150         }
 151         srcu_read_unlock(&kvm->srcu, idx);
 152
 153         return r;
 154 }
 155
 156 int kvm_assign_device(struct kvm *kvm,
 157                       struct kvm_assigned_dev_kernel *assigned_dev)
 158 {
 159         struct pci_dev *pdev = NULL;
 160         struct iommu_domain *domain = kvm->arch.iommu_domain;
 161         int r;
 162         bool noncoherent;
 163
 164         /* check if iommu exists and in use */
 165         if (!domain)
 166                 return 0;
 167
 168         pdev = assigned_dev->dev;
 169         if (pdev == NULL)
 170                 return -ENODEV;
 171
 172         r = iommu_attach_device(domain, &pdev->dev);
 173         if (r) {
 174                 dev_err(&pdev->dev, "kvm assign device failed ret %d", r);
 175                 return r;
 176         }
 177
 178         noncoherent = !iommu_domain_has_cap(kvm->arch.iommu_domain,
 179                                             IOMMU_CAP_CACHE_COHERENCY);
 180
 181         /* Check if need to update IOMMU page table for guest memory */
 182         if (noncoherent != kvm->arch.iommu_noncoherent) {
 183                 kvm_iommu_unmap_memslots(kvm);
 184                 kvm->arch.iommu_noncoherent = noncoherent;
 185                 r = kvm_iommu_map_memslots(kvm);
 186                 if (r)
 187                         goto out_unmap;
 188         }
 189
 190         pdev->dev_flags |= PCI_DEV_FLAGS_ASSIGNED;
 191
 192         dev_info(&pdev->dev, "kvm assign device\n");
 193
 194         return 0;
 195 out_unmap:
 196         kvm_iommu_unmap_memslots(kvm);
 197         return r;
 198 }
 199
 200 int kvm_deassign_device(struct kvm *kvm,
 201                         struct kvm_assigned_dev_kernel *assigned_dev)
 202 {
 203         struct iommu_domain *domain = kvm->arch.iommu_domain;
 204         struct pci_dev *pdev = NULL;
 205
 206         /* check if iommu exists and in use */
 207         if (!domain)
 208                 return 0;
 209
 210         pdev = assigned_dev->dev;
 211         if (pdev == NULL)
 212                 return -ENODEV;
 213
 214         iommu_detach_device(domain, &pdev->dev);
 215
 216         pdev->dev_flags &= ~PCI_DEV_FLAGS_ASSIGNED;
 217
 218         dev_info(&pdev->dev, "kvm deassign device\n");
 219
 220         return 0;
 221 }
 222
 223 int kvm_iommu_map_guest(struct kvm *kvm)
 224 {
 225         int r;
 226
 227         if (!iommu_present(&pci_bus_type)) {
 228                 printk(KERN_ERR "%s: iommu not found\n", __func__);
 229                 return -ENODEV;
 230         }
 231
 232         mutex_lock(&kvm->slots_lock);
 233
 234         kvm->arch.iommu_domain = iommu_domain_alloc(&pci_bus_type);
 235         if (!kvm->arch.iommu_domain) {
 236                 r = -ENOMEM;
 237                 goto out_unlock;
 238         }
 239
 240         if (!allow_unsafe_assigned_interrupts &&
 241             !iommu_domain_has_cap(kvm->arch.iommu_domain,
 242                                   IOMMU_CAP_INTR_REMAP)) {
 243                 printk(KERN_WARNING "%s: No interrupt remapping support,"
 244                        " disallowing device assignment."
 245                        " Re-enble with \"allow_unsafe_assigned_interrupts=1\""
 246                        " module option.\n", __func__);
 247                 iommu_domain_free(kvm->arch.iommu_domain);
 248                 kvm->arch.iommu_domain = NULL;
 249                 r = -EPERM;
 250                 goto out_unlock;
 251         }
 252
 253         r = kvm_iommu_map_memslots(kvm);
 254         if (r)
 255                 kvm_iommu_unmap_memslots(kvm);
 256
 257 out_unlock:
 258         mutex_unlock(&kvm->slots_lock);
 259         return r;
 260 }
 261
 262 static void kvm_unpin_pages(struct kvm *kvm, pfn_t pfn, unsigned long npages)
 263 {
 264         unsigned long i;
 265
 266         for (i = 0; i < npages; ++i)
 267                 kvm_release_pfn_clean(pfn + i);
 268 }
 269
 270 static void kvm_iommu_put_pages(struct kvm *kvm,
 271                                 gfn_t base_gfn, unsigned long npages)
 272 {
 273         struct iommu_domain *domain;
 274         gfn_t end_gfn, gfn;
 275         pfn_t pfn;
 276         u64 phys;
 277
 278         domain  = kvm->arch.iommu_domain;
 279         end_gfn = base_gfn + npages;
 280         gfn     = base_gfn;
 281
 282         /* check if iommu exists and in use */
 283         if (!domain)
 284                 return;
 285
 286         while (gfn < end_gfn) {
 287                 unsigned long unmap_pages;
 288                 size_t size;
 289
 290                 /* Get physical address */
 291                 phys = iommu_iova_to_phys(domain, gfn_to_gpa(gfn));
 292
 293                 if (!phys) {
 294                         gfn++;
 295                         continue;
 296                 }
 297
 298                 pfn  = phys >> PAGE_SHIFT;
 299
 300                 /* Unmap address from IO address space */
 301                 size       = iommu_unmap(domain, gfn_to_gpa(gfn), PAGE_SIZE);
 302                 unmap_pages = 1ULL << get_order(size);
 303
 304                 /* Unpin all pages we just unmapped to not leak any memory */
 305                 kvm_unpin_pages(kvm, pfn, unmap_pages);
 306
 307                 gfn += unmap_pages;
 308         }
 309 }
 310
 311 void kvm_iommu_unmap_pages(struct kvm *kvm, struct kvm_memory_slot *slot)
 312 {
 313         kvm_iommu_put_pages(kvm, slot->base_gfn, slot->npages);
 314 }
 315
 316 static int kvm_iommu_unmap_memslots(struct kvm *kvm)
 317 {
 318         int idx;
 319         struct kvm_memslots *slots;
 320         struct kvm_memory_slot *memslot;
 321
 322         idx = srcu_read_lock(&kvm->srcu);
 323         slots = kvm_memslots(kvm);
 324
 325         kvm_for_each_memslot(memslot, slots)
 326                 kvm_iommu_unmap_pages(kvm, memslot);
 327
 328         srcu_read_unlock(&kvm->srcu, idx);
 329
 330         return 0;
 331 }
 332
 333 int kvm_iommu_unmap_guest(struct kvm *kvm)
 334 {
 335         struct iommu_domain *domain = kvm->arch.iommu_domain;
 336
 337         /* check if iommu exists and in use */
 338         if (!domain)
 339                 return 0;
 340
 341         mutex_lock(&kvm->slots_lock);
 342         kvm_iommu_unmap_memslots(kvm);
 343         kvm->arch.iommu_domain = NULL;
 344         kvm->arch.iommu_noncoherent = false;
 345         mutex_unlock(&kvm->slots_lock);
 346
 347         iommu_domain_free(domain);
 348         return 0;
 349 }